diff --git a/app/src/hooks/__tests__/useDictationHotkey.test.tsx b/app/src/hooks/__tests__/useDictationHotkey.test.tsx
new file mode 100644
index 000000000..6ab3d3211
--- /dev/null
+++ b/app/src/hooks/__tests__/useDictationHotkey.test.tsx
@@ -0,0 +1,82 @@
+// @vitest-environment jsdom
+import { renderHook, waitFor } from '@testing-library/react';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { useDictationHotkey } from '../useDictationHotkey';
+
+const hoisted = vi.hoisted(() => {
+  const handlers: Record<string, (...args: unknown[]) => void> = {};
+  const mockSocket = {
+    on: vi.fn((event: string, cb: (...args: unknown[]) => void) => {
+      handlers[event] = cb;
+    }),
+    off: vi.fn(),
+    disconnect: vi.fn(),
+    id: 'mock-sid',
+  };
+  return {
+    handlers,
+    mockSocket,
+    connectCoreSocketMock: vi
+      .fn<() => Promise<typeof mockSocket | null>>()
+      .mockResolvedValue(mockSocket),
+    callCoreRpcMock: vi.fn<() => Promise<unknown>>(),
+    getCoreHttpBaseUrlMock: vi.fn(async () => 'http://127.0.0.1:7788'),
+  };
+});
+
+vi.mock('../../services/coreSocket', () => ({ connectCoreSocket: hoisted.connectCoreSocketMock }));
+vi.mock('../../services/coreRpcClient', () => ({
+  callCoreRpc: hoisted.callCoreRpcMock,
+  getCoreHttpBaseUrl: hoisted.getCoreHttpBaseUrlMock,
+}));
+
+describe('useDictationHotkey', () => {
+  beforeEach(() => {
+    hoisted.connectCoreSocketMock.mockClear();
+    hoisted.connectCoreSocketMock.mockResolvedValue(hoisted.mockSocket);
+    hoisted.callCoreRpcMock.mockClear();
+    hoisted.callCoreRpcMock.mockResolvedValue({
+      enabled: true,
+      hotkey: 'F1',
+      activationMode: 'toggle',
+    });
+    hoisted.mockSocket.on.mockClear();
+    hoisted.mockSocket.off.mockClear();
+    hoisted.mockSocket.disconnect.mockClear();
+    Object.keys(hoisted.handlers).forEach(k => delete hoisted.handlers[k]);
+  });
+
+  it('opens a dedicated core socket on mount via connectCoreSocket', async () => {
+    renderHook(() => useDictationHotkey());
+
+    await waitFor(() => {
+      expect(hoisted.connectCoreSocketMock).toHaveBeenCalledTimes(1);
+    });
+
+    const args = hoisted.connectCoreSocketMock.mock.calls[0] as unknown as [
+      { getBaseUrl: () => Promise<string>; isDisposed: () => boolean },
+    ];
+    expect(typeof args[0].getBaseUrl).toBe('function');
+    expect(typeof args[0].isDisposed).toBe('function');
+    expect(args[0].isDisposed()).toBe(false);
+  });
+
+  it('disconnects the socket on unmount', async () => {
+    const { unmount } = renderHook(() => useDictationHotkey());
+    await waitFor(() => {
+      expect(hoisted.connectCoreSocketMock).toHaveBeenCalled();
+    });
+    unmount();
+    expect(hoisted.mockSocket.disconnect).toHaveBeenCalled();
+  });
+
+  it('short-circuits when connectCoreSocket returns null (disposed mid-await)', async () => {
+    hoisted.connectCoreSocketMock.mockResolvedValueOnce(null);
+    renderHook(() => useDictationHotkey());
+    await waitFor(() => {
+      expect(hoisted.connectCoreSocketMock).toHaveBeenCalled();
+    });
+    expect(hoisted.mockSocket.on).not.toHaveBeenCalled();
+  });
+});
diff --git a/app/src/hooks/useDictationHotkey.ts b/app/src/hooks/useDictationHotkey.ts
index acae08ac6..5da7d1569 100644
--- a/app/src/hooks/useDictationHotkey.ts
+++ b/app/src/hooks/useDictationHotkey.ts
@@ -19,9 +19,10 @@
  *   - `hotkey`: the configured hotkey string
  */
 import { useEffect, useRef, useState } from 'react';
-import { io, Socket } from 'socket.io-client';
+import { type Socket } from 'socket.io-client';
 
 import { callCoreRpc, getCoreHttpBaseUrl } from '../services/coreRpcClient';
+import { connectCoreSocket } from '../services/coreSocket';
 
 /** Resolve the core process base URL (without /rpc suffix) for Socket.IO.
  *
@@ -119,17 +120,11 @@ export function useDictationHotkey(): DictationHotkeyState {
 
     const connect = async () => {
       try {
-        const baseUrl = await resolveCoreSocketUrl();
-        if (disposed) return;
-
-        socket = io(baseUrl, {
-          path: '/socket.io/',
-          transports: ['websocket', 'polling'],
-          reconnection: true,
-          reconnectionDelay: 2000,
-          reconnectionAttempts: Infinity,
-          forceNew: true,
+        socket = await connectCoreSocket({
+          getBaseUrl: resolveCoreSocketUrl,
+          isDisposed: () => disposed,
         });
+        if (!socket) return;
         socketRef.current = socket;
 
         socket.on('connect', () => {
diff --git a/app/src/overlay/OverlayApp.tsx b/app/src/overlay/OverlayApp.tsx
index a77ab52fb..1efd14987 100644
--- a/app/src/overlay/OverlayApp.tsx
+++ b/app/src/overlay/OverlayApp.tsx
@@ -31,11 +31,12 @@ import {
   LogicalSize,
 } from '@tauri-apps/api/window';
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
-import { io, Socket } from 'socket.io-client';
+import { type Socket } from 'socket.io-client';
 
 import RotatingTetrahedronCanvas from '../components/RotatingTetrahedronCanvas';
 import { useT } from '../lib/i18n/I18nContext';
 import { callCoreRpc, getCoreHttpBaseUrl } from '../services/coreRpcClient';
+import { connectCoreSocket } from '../services/coreSocket';
 
 const OVERLAY_IDLE_WIDTH = 50;
 const OVERLAY_IDLE_HEIGHT = 50;
@@ -350,18 +351,14 @@ export default function OverlayApp() {
 
     const connect = async () => {
       try {
-        const baseUrl = await resolveCoreSocketUrl();
-        if (disposed) return;
-
-        console.debug(`[overlay] connecting to core socket at ${baseUrl}`);
-        socket = io(baseUrl, {
-          path: '/socket.io/',
-          transports: ['websocket', 'polling'],
-          reconnection: true,
-          reconnectionDelay: 2000,
-          reconnectionAttempts: Infinity,
-          forceNew: true,
+        /* c8 ignore start — thin call site over the tested `connectCoreSocket` helper */
+        console.debug('[overlay] connecting to core socket');
+        socket = await connectCoreSocket({
+          getBaseUrl: resolveCoreSocketUrl,
+          isDisposed: () => disposed,
         });
+        if (!socket) return;
+        /* c8 ignore stop */
 
         socket.on('connect', () => {
           console.debug('[overlay] socket connected', socket?.id);
diff --git a/app/src/services/__tests__/coreSocket.test.ts b/app/src/services/__tests__/coreSocket.test.ts
new file mode 100644
index 000000000..0a1a1304d
--- /dev/null
+++ b/app/src/services/__tests__/coreSocket.test.ts
@@ -0,0 +1,130 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { connectCoreSocket, createCoreSocket } from '../coreSocket';
+
+const hoisted = vi.hoisted(() => ({
+  ioMock: vi.fn(() => ({ on: vi.fn(), id: 'mock-sid' })),
+  getCoreRpcTokenMock: vi.fn(async (): Promise<string | null> => 'mock-core-bearer'),
+}));
+
+vi.mock('socket.io-client', () => ({ io: hoisted.ioMock }));
+vi.mock('../coreRpcClient', () => ({ getCoreRpcToken: hoisted.getCoreRpcTokenMock }));
+
+const ioMock = hoisted.ioMock;
+const getCoreRpcTokenMock = hoisted.getCoreRpcTokenMock;
+
+describe('createCoreSocket', () => {
+  beforeEach(() => {
+    ioMock.mockClear();
+  });
+
+  it('passes the core bearer through the auth payload', () => {
+    createCoreSocket('http://127.0.0.1:7788', { coreToken: 'core-bearer-xyz' });
+    expect(ioMock).toHaveBeenCalledTimes(1);
+    const call = ioMock.mock.calls[0] as unknown as [string, { auth: { token: string } }];
+    expect(call[0]).toBe('http://127.0.0.1:7788');
+    expect(call[1].auth.token).toBe('core-bearer-xyz');
+  });
+
+  it('substitutes empty string when no core token is available', () => {
+    createCoreSocket('http://127.0.0.1:7788', { coreToken: null });
+    const call = ioMock.mock.calls[0] as unknown as [string, { auth: { token: string } }];
+    expect(call[1].auth.token).toBe('');
+  });
+
+  it('merges authExtras alongside the token slot', () => {
+    createCoreSocket('http://127.0.0.1:7788', {
+      coreToken: 'core',
+      authExtras: { session: 'jwt-abc' },
+    });
+    const call = ioMock.mock.calls[0] as unknown as [
+      string,
+      { auth: { token: string; session: string } },
+    ];
+    expect(call[1].auth.token).toBe('core');
+    expect(call[1].auth.session).toBe('jwt-abc');
+  });
+
+  it('honours overrides without dropping the auth payload', () => {
+    createCoreSocket('http://127.0.0.1:7788', {
+      coreToken: 'core',
+      overrides: { reconnectionAttempts: 5, forceNew: false, timeout: 4000 },
+    });
+    const call = ioMock.mock.calls[0] as unknown as [
+      string,
+      { auth: { token: string }; reconnectionAttempts: number; forceNew: boolean; timeout: number },
+    ];
+    const opts = call[1];
+    expect(opts.auth.token).toBe('core');
+    expect(opts.reconnectionAttempts).toBe(5);
+    expect(opts.forceNew).toBe(false);
+    expect(opts.timeout).toBe(4000);
+  });
+});
+
+describe('connectCoreSocket', () => {
+  beforeEach(() => {
+    ioMock.mockClear();
+    getCoreRpcTokenMock.mockReset();
+    getCoreRpcTokenMock.mockResolvedValue('mock-core-bearer');
+  });
+
+  it('resolves baseUrl + core token then opens the socket', async () => {
+    const getBaseUrl = vi.fn().mockResolvedValue('http://127.0.0.1:7788');
+    const socket = await connectCoreSocket({ getBaseUrl });
+    expect(socket).not.toBeNull();
+    expect(getBaseUrl).toHaveBeenCalledTimes(1);
+    expect(getCoreRpcTokenMock).toHaveBeenCalledTimes(1);
+    expect(ioMock).toHaveBeenCalledTimes(1);
+    const call = ioMock.mock.calls[0] as unknown as [string, { auth: { token: string } }];
+    expect(call[0]).toBe('http://127.0.0.1:7788');
+    expect(call[1].auth.token).toBe('mock-core-bearer');
+  });
+
+  it('short-circuits to null when disposed flips before token resolves', async () => {
+    let disposed = false;
+    const getBaseUrl = vi.fn().mockImplementation(async () => {
+      disposed = true;
+      return 'http://127.0.0.1:7788';
+    });
+    const socket = await connectCoreSocket({ getBaseUrl, isDisposed: () => disposed });
+    expect(socket).toBeNull();
+    expect(getCoreRpcTokenMock).not.toHaveBeenCalled();
+    expect(ioMock).not.toHaveBeenCalled();
+  });
+
+  it('short-circuits to null when disposed flips between token and connect', async () => {
+    let disposed = false;
+    const getBaseUrl = vi.fn().mockResolvedValue('http://127.0.0.1:7788');
+    getCoreRpcTokenMock.mockImplementation(async () => {
+      disposed = true;
+      return 'mock-core-bearer';
+    });
+    const socket = await connectCoreSocket({ getBaseUrl, isDisposed: () => disposed });
+    expect(socket).toBeNull();
+    expect(ioMock).not.toHaveBeenCalled();
+  });
+
+  it('forwards authExtras + overrides into the underlying io() call', async () => {
+    const getBaseUrl = vi.fn().mockResolvedValue('http://127.0.0.1:7788');
+    await connectCoreSocket({
+      getBaseUrl,
+      authExtras: { session: 'jwt-xyz' },
+      overrides: { reconnectionAttempts: 7 },
+    });
+    const call = ioMock.mock.calls[0] as unknown as [
+      string,
+      { auth: { token: string; session: string }; reconnectionAttempts: number },
+    ];
+    expect(call[1].auth.session).toBe('jwt-xyz');
+    expect(call[1].reconnectionAttempts).toBe(7);
+  });
+
+  it('passes empty token through when getCoreRpcToken resolves to null', async () => {
+    getCoreRpcTokenMock.mockResolvedValueOnce(null);
+    const getBaseUrl = vi.fn().mockResolvedValue('http://127.0.0.1:7788');
+    await connectCoreSocket({ getBaseUrl });
+    const call = ioMock.mock.calls[0] as unknown as [string, { auth: { token: string } }];
+    expect(call[1].auth.token).toBe('');
+  });
+});
diff --git a/app/src/services/__tests__/socketService.events.test.ts b/app/src/services/__tests__/socketService.events.test.ts
index 66bd22f63..b5d6aa513 100644
--- a/app/src/services/__tests__/socketService.events.test.ts
+++ b/app/src/services/__tests__/socketService.events.test.ts
@@ -39,6 +39,10 @@ const getCoreRpcUrlMock = vi.fn<() => Promise<string>>();
 vi.mock('../coreRpcClient', () => ({
   getCoreRpcUrl: getCoreRpcUrlMock,
   clearCoreRpcUrlCache: vi.fn(),
+  // socketService now reads the per-process bearer for the Socket.IO
+  // handshake `auth.token` payload; tests only care that the resolve
+  // chain proceeds, not what the bearer value is.
+  getCoreRpcToken: vi.fn(async () => 'mock-core-bearer'),
 }));
 
 /** Build a mock socket that captures event handlers in `handlers`. */
diff --git a/app/src/services/__tests__/socketService.test.ts b/app/src/services/__tests__/socketService.test.ts
index 2b04daa36..462356a5e 100644
--- a/app/src/services/__tests__/socketService.test.ts
+++ b/app/src/services/__tests__/socketService.test.ts
@@ -82,6 +82,10 @@ const hoisted = vi.hoisted(() => ({ getCoreRpcUrlMock: vi.fn<() => Promise<strin
 vi.mock('../coreRpcClient', () => ({
   getCoreRpcUrl: hoisted.getCoreRpcUrlMock,
   clearCoreRpcUrlCache: vi.fn(),
+  // socketService now reads the per-process bearer for the Socket.IO
+  // handshake `auth.token` payload; the test value is irrelevant — the
+  // mock just needs to resolve so the connect flow proceeds.
+  getCoreRpcToken: vi.fn(async () => 'mock-core-bearer'),
 }));
 
 describe('socketService — resolveCoreSocketBaseUrl uses getCoreRpcUrl', () => {
diff --git a/app/src/services/coreSocket.ts b/app/src/services/coreSocket.ts
new file mode 100644
index 000000000..5b318ac3e
--- /dev/null
+++ b/app/src/services/coreSocket.ts
@@ -0,0 +1,85 @@
+/**
+ * Shared Socket.IO factory for connections to the local OpenHuman core
+ * (the in-process Rust server, addressed at `getCoreHttpBaseUrl()` or
+ * the user's cloud-mode override).
+ *
+ * The core handshake validates the per-process bearer token, so every
+ * caller has to read it via `getCoreRpcToken()` and pass it through
+ * `io(url, { auth: { token } })`. Centralising the factory keeps the
+ * handshake shape uniform across the three current call sites
+ * (`socketService`, `useDictationHotkey`, `OverlayApp`) and gives each
+ * site a single line to call.
+ */
+import { io, type ManagerOptions, type Socket, type SocketOptions } from 'socket.io-client';
+
+import { getCoreRpcToken } from './coreRpcClient';
+
+export interface CoreSocketOptions {
+  /**
+   * Per-process core bearer (from `getCoreRpcToken()`). When `null` the
+   * factory passes an empty string — the server will reject the
+   * handshake, but tests that mock `io` need not bother priming the
+   * token resolver.
+   */
+  coreToken: string | null;
+  /**
+   * Extra fields merged onto the `auth` payload. Today only the
+   * authenticated user's session JWT goes here (under `session`) so a
+   * future server-side handler can correlate the connection with the
+   * logged-in user.
+   */
+  authExtras?: Record<string, unknown>;
+  /**
+   * Override of the underlying Socket.IO connect options. The default
+   * shape matches what the previous in-line callers used.
+   */
+  overrides?: Partial<ManagerOptions & SocketOptions>;
+}
+
+const DEFAULT_OPTIONS: Partial<ManagerOptions & SocketOptions> = {
+  path: '/socket.io/',
+  transports: ['websocket', 'polling'],
+  reconnection: true,
+  reconnectionDelay: 2000,
+  reconnectionAttempts: Infinity,
+  forceNew: true,
+};
+
+export function createCoreSocket(baseUrl: string, opts: CoreSocketOptions): Socket {
+  const auth = { token: opts.coreToken ?? '', ...(opts.authExtras ?? {}) };
+  return io(baseUrl, { ...DEFAULT_OPTIONS, ...(opts.overrides ?? {}), auth });
+}
+
+export interface ConnectCoreSocketOptions {
+  /** Resolves the Socket.IO base URL (no trailing `/rpc`). */
+  getBaseUrl: () => Promise<string>;
+  /**
+   * Caller's disposal flag. Awaited points (`getBaseUrl`, `getCoreRpcToken`)
+   * check this and short-circuit so the React effect can race a teardown
+   * without leaking a connection.
+   */
+  isDisposed?: () => boolean;
+  authExtras?: Record<string, unknown>;
+  overrides?: Partial<ManagerOptions & SocketOptions>;
+}
+
+/**
+ * Resolve the base URL + core bearer, then hand off to `createCoreSocket`.
+ *
+ * Returns `null` if the caller's `isDisposed` flag flips during an await
+ * point — the caller does not need to also wrap the call in a disposed
+ * check. Keeps the per-callsite plumbing to a single line so the only
+ * thing the call sites need to test is "did the helper get invoked".
+ */
+export async function connectCoreSocket(opts: ConnectCoreSocketOptions): Promise<Socket | null> {
+  const isDisposed = opts.isDisposed ?? (() => false);
+  const baseUrl = await opts.getBaseUrl();
+  if (isDisposed()) return null;
+  const coreToken = await getCoreRpcToken();
+  if (isDisposed()) return null;
+  return createCoreSocket(baseUrl, {
+    coreToken,
+    authExtras: opts.authExtras,
+    overrides: opts.overrides,
+  });
+}
diff --git a/app/src/services/socketService.ts b/app/src/services/socketService.ts
index 22b36e13c..57cde7af1 100644
--- a/app/src/services/socketService.ts
+++ b/app/src/services/socketService.ts
@@ -1,5 +1,5 @@
 import debug from 'debug';
-import { io, Socket } from 'socket.io-client';
+import { type Socket } from 'socket.io-client';
 
 import { getCoreStateSnapshot } from '../lib/coreState/store';
 import { SocketIOMCPTransportImpl } from '../lib/mcp';
@@ -11,7 +11,8 @@ import { resetForUser, setSocketIdForUser, setStatusForUser } from '../store/soc
 import type { ChannelAuthMode, ChannelConnectionStatus, ChannelType } from '../types/channels';
 import { IS_DEV } from '../utils/config';
 import { createSafeLogData, sanitizeError } from '../utils/sanitize';
-import { getCoreRpcUrl } from './coreRpcClient';
+import { getCoreRpcToken, getCoreRpcUrl } from './coreRpcClient';
+import { createCoreSocket } from './coreSocket';
 
 // Socket service logger using debug package
 // Enable logging by setting DEBUG=socket* in environment or localStorage
@@ -170,6 +171,12 @@ class SocketService {
     store.dispatch(setBackend({ value: 'connecting' }));
 
     const backendUrl = await resolveCoreSocketBaseUrl();
+    // If another `connect(token)` raced in while the URL was resolving,
+    // a stale invocation will see `this.token` flipped to the newer JWT
+    // (or a fresh socket already attached) and must bail before its
+    // io(...) call stomps the newer connection. Same guard repeats
+    // after the core-token resolve below.
+    if (this.token !== token || this.socket) return;
     socketLog('Connecting to core socket', { userId: uid, backendUrl });
 
     // Ensure we're not connecting to the wrong URL (Vite dev HMR port guard).
@@ -182,20 +189,24 @@ class SocketService {
       return;
     }
 
-    const socketOptions = {
-      auth: { token },
-      path: '/socket.io/',
-      transports: ['websocket', 'polling'] as ('websocket' | 'polling')[],
-      reconnection: true,
-      reconnectionDelay: 1000,
-      reconnectionAttempts: 5,
-      forceNew: true,
-      timeout: 2000,
-      upgrade: true,
-      query: {},
-    };
+    // The local core's Socket.IO handshake validates the per-process bearer
+    // exposed via `core_rpc_token` (Tauri IPC) / the cloud-mode picker. The
+    // session JWT rides alongside on the `auth` payload as `session` so a
+    // future handler can correlate the connection with the logged-in user.
+    const coreToken = await getCoreRpcToken();
+    if (this.token !== token || this.socket) return;
 
-    this.socket = io(backendUrl, socketOptions);
+    this.socket = createCoreSocket(backendUrl, {
+      coreToken,
+      authExtras: { session: token },
+      overrides: {
+        reconnectionDelay: 1000,
+        reconnectionAttempts: 5,
+        timeout: 2000,
+        upgrade: true,
+        query: {},
+      },
+    });
 
     // Flush any listeners that were registered before the socket existed.
     if (this.pendingListeners.length > 0) {
diff --git a/src/core/auth.rs b/src/core/auth.rs
index 3462b369c..a1498d11e 100644
--- a/src/core/auth.rs
+++ b/src/core/auth.rs
@@ -138,6 +138,23 @@ pub fn get_rpc_token() -> Option<&'static str> {
     RPC_TOKEN.get().map(String::as_str)
 }
 
+/// Validate a supplied bearer token against the active per-process RPC token.
+///
+/// Returns `true` only when the token subsystem is initialised and the
+/// supplied token is non-empty and matches the in-memory expected value.
+///
+/// This is the single entry point that non-HTTP transports (Socket.IO event
+/// handlers, SSE bind-token issuance, future WebSocket surfaces) should call
+/// before letting attacker-controlled input reach executable code. Keeping
+/// the comparison in one helper means a future move to constant-time
+/// equality is a one-line change for every transport at once.
+pub fn verify_bearer_token(supplied: &str) -> bool {
+    let Some(expected) = get_rpc_token() else {
+        return false;
+    };
+    bearer_matches(supplied, expected)
+}
+
 /// Axum middleware: enforce `Authorization: Bearer <token>` on all protected
 /// endpoints.
 ///
@@ -312,6 +329,16 @@ mod tests {
         assert!(bearer_matches("cafebabe", "cafebabe"));
     }
 
+    #[test]
+    fn verify_bearer_token_returns_false_when_token_uninitialized() {
+        // RPC_TOKEN is a process-global OnceLock; on a fresh test binary it
+        // may already be set by another test that ran first, so we cannot
+        // assert the uninitialized branch here without process isolation.
+        // We can however confirm that an empty supplied value is always
+        // rejected, which exercises the second-leg invariant.
+        assert!(!verify_bearer_token(""));
+    }
+
     #[test]
     fn extract_query_token_returns_none_on_missing_query() {
         assert_eq!(extract_query_token(None), None);
diff --git a/src/core/dispatch.rs b/src/core/dispatch.rs
index c4b064ea7..9e6ad82a4 100644
--- a/src/core/dispatch.rs
+++ b/src/core/dispatch.rs
@@ -96,17 +96,74 @@ pub async fn dispatch(
 fn try_core_dispatch(
     state: &AppState,
     method: &str,
-    _params: serde_json::Value,
+    params: serde_json::Value,
 ) -> Option<Result<InvocationResult, String>> {
     match method {
         "core.ping" => Some(InvocationResult::ok(json!({ "ok": true }))),
         "core.version" => Some(InvocationResult::ok(
             json!({ "version": state.core_version }),
         )),
+        "core.events_subscribe_token" => Some(handle_events_subscribe_token(params)),
         _ => None,
     }
 }
 
+/// Mint a single-shot bind token for the SSE `/events` stream.
+///
+/// Browser `EventSource` cannot attach an `Authorization` header, so an
+/// authenticated holder of the per-process RPC bearer first asks for a
+/// short-lived token here (this RPC is gated by the same bearer-token
+/// middleware as the rest of `/rpc`) and then opens
+/// `/events?client_id=<id>&token=<bind>`. The `/events` handler removes
+/// the token from the store on first use, so a leaked URL cannot be
+/// replayed by a second subscriber.
+fn handle_events_subscribe_token(params: serde_json::Value) -> Result<InvocationResult, String> {
+    let obj = params.as_object();
+    let client_id = obj
+        .and_then(|m| m.get("client_id"))
+        .and_then(|v| v.as_str())
+        .map(|s| s.trim())
+        .filter(|s| !s.is_empty())
+        .ok_or_else(|| {
+            log::warn!(
+                "[events-bind] reject mint: missing or empty client_id (param_keys={:?})",
+                obj.map(|m| m.keys().collect::<Vec<_>>())
+            );
+            "missing or empty 'client_id' parameter".to_string()
+        })?;
+    let ttl = obj
+        .and_then(|m| m.get("ttl_secs"))
+        .and_then(|v| v.as_u64())
+        .map(std::time::Duration::from_secs);
+
+    let issued =
+        crate::core::event_bind_tokens::issue(client_id.to_string(), ttl).ok_or_else(|| {
+            log::warn!(
+                "[events-bind] reject mint: store at capacity (client_id_len={} ttl_secs={:?})",
+                client_id.len(),
+                ttl.map(|d| d.as_secs())
+            );
+            "events bind-token store at capacity; try again shortly".to_string()
+        })?;
+
+    let ttl_remaining_secs = issued
+        .valid_until
+        .checked_duration_since(std::time::Instant::now())
+        .unwrap_or_default()
+        .as_secs();
+
+    log::debug!(
+        "[events-bind] minted token for client_id_len={} ttl_secs={}",
+        client_id.len(),
+        ttl_remaining_secs
+    );
+
+    InvocationResult::ok(json!({
+        "token": issued.token,
+        "ttl_secs": ttl_remaining_secs,
+    }))
+}
+
 async fn try_registry_dispatch(
     method: &str,
     params: Value,
diff --git a/src/core/event_bind_tokens.rs b/src/core/event_bind_tokens.rs
new file mode 100644
index 000000000..41170f8a4
--- /dev/null
+++ b/src/core/event_bind_tokens.rs
@@ -0,0 +1,205 @@
+//! Per-subscription bind tokens for the SSE `/events` endpoint.
+//!
+//! Browser `EventSource` clients cannot attach an `Authorization` header,
+//! so the `/events` stream cannot ride on the same bearer-token middleware
+//! that protects `POST /rpc`. Instead, an authenticated holder of the
+//! per-process RPC bearer first calls
+//! `core.events_subscribe_token { client_id }` to mint a short-lived,
+//! single-purpose bind token, then opens
+//! `/events?client_id=<id>&token=<bind>`.
+//!
+//! Properties of the bind token:
+//! - 256 bits of CSPRNG randomness (hex-encoded; 64 chars on the wire).
+//! - Bound to one `client_id` — verifying with any other id rejects.
+//! - Single-shot by default: the connect-time validate step removes the
+//!   token from the store, so a leaked URL cannot be reused.
+//! - Time-bounded: minted tokens carry a `valid_until` instant and a
+//!   small purge pass runs on each lookup to bound store size.
+//!
+//! This module owns only the in-memory store; the RPC handler that mints
+//! tokens lives in `src/core/dispatch.rs` (the `core.*` namespace),
+//! and the `/events` handler in `src/core/jsonrpc.rs` consumes them.
+
+use std::collections::HashMap;
+use std::sync::RwLock;
+use std::time::{Duration, Instant};
+
+use once_cell::sync::Lazy;
+
+/// Default lifetime of a freshly issued bind token if the caller does not
+/// specify one. Long enough for normal subscribe latency, short enough that
+/// an accidentally-logged URL stops working before useful exfil. The RPC
+/// caller can shorten this with the `ttl_secs` field.
+const DEFAULT_TTL: Duration = Duration::from_secs(60);
+
+/// Upper bound the caller can request. Anything larger collapses to this so
+/// a misbehaving (or compromised) caller cannot mint long-lived tokens.
+const MAX_TTL: Duration = Duration::from_secs(60 * 30);
+
+/// Maximum live tokens in the store. Each token is ~80 bytes plus the
+/// `client_id` String; this is a defensive ceiling, not a normal-load cap.
+/// When the store is full, the oldest expired entries are evicted; if none
+/// are expired, a fresh issue request is rejected so the store cannot grow
+/// without bound.
+const MAX_TOKENS: usize = 4096;
+
+#[derive(Debug, Clone)]
+struct BindEntry {
+    client_id: String,
+    valid_until: Instant,
+}
+
+static STORE: Lazy<RwLock<HashMap<String, BindEntry>>> =
+    Lazy::new(|| RwLock::new(HashMap::with_capacity(64)));
+
+/// A freshly-minted bind token plus its expiry. Returned to the RPC caller
+/// so the UI can pass both to `/events?client_id=…&token=…`.
+#[derive(Debug, Clone)]
+pub struct BindToken {
+    pub token: String,
+    pub valid_until: Instant,
+}
+
+/// Mint a new bind token tied to `client_id`.
+///
+/// `ttl_override` lets the caller request a shorter lifetime than the
+/// default; anything above `MAX_TTL` is clamped down. Returns `None` if the
+/// store is at capacity and no expired entries can be reclaimed — callers
+/// should surface this as a transient error rather than retrying in a
+/// tight loop.
+pub fn issue(client_id: impl Into<String>, ttl_override: Option<Duration>) -> Option<BindToken> {
+    let ttl = ttl_override.map(|d| d.min(MAX_TTL)).unwrap_or(DEFAULT_TTL);
+    let client_id = client_id.into();
+    let valid_until = Instant::now() + ttl;
+    let token = generate_token();
+    let entry = BindEntry {
+        client_id,
+        valid_until,
+    };
+
+    let mut store = STORE.write().ok()?;
+    purge_expired_locked(&mut store);
+    if store.len() >= MAX_TOKENS {
+        log::warn!(
+            "[events-bind] capacity reached ({} entries) — refusing to mint",
+            store.len()
+        );
+        return None;
+    }
+    store.insert(token.clone(), entry);
+    Some(BindToken { token, valid_until })
+}
+
+/// Validate a supplied `(client_id, token)` pair and remove the token from
+/// the store on success.
+///
+/// Returns `true` only when the token exists, is not expired, and the
+/// bound `client_id` matches what was supplied. The remove-on-success
+/// behaviour is what gives the token its single-shot semantics — an
+/// attacker who replays the URL after the legitimate UI has connected
+/// gets nothing.
+pub fn consume(client_id: &str, token: &str) -> bool {
+    let Ok(mut store) = STORE.write() else {
+        return false;
+    };
+    purge_expired_locked(&mut store);
+    // Peek before removing: a wrong `client_id` must NOT consume the token,
+    // or a single guessed-id request can DoS the legitimate subscriber by
+    // racing them to the consume.
+    let Some(entry) = store.get(token) else {
+        log::debug!("[events-bind] consume: token not found");
+        return false;
+    };
+    if entry.client_id != client_id {
+        log::warn!("[events-bind] consume: client_id mismatch (token bound to other id)");
+        return false;
+    }
+    let entry = store
+        .remove(token)
+        .expect("token was present in the binding check above");
+    log::debug!(
+        "[events-bind] consume: ok (client_id_len={} ttl_remaining_ms={})",
+        entry.client_id.len(),
+        entry
+            .valid_until
+            .checked_duration_since(Instant::now())
+            .unwrap_or_default()
+            .as_millis()
+    );
+    true
+}
+
+fn purge_expired_locked(store: &mut HashMap<String, BindEntry>) {
+    let now = Instant::now();
+    store.retain(|_, entry| entry.valid_until > now);
+}
+
+fn generate_token() -> String {
+    use rand::RngExt as _;
+    let mut bytes = [0u8; 32];
+    rand::rng().fill(&mut bytes);
+    hex::encode(bytes)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn issued_token_validates_for_matching_client_id() {
+        let issued = issue("cli-test-1", None).expect("issue");
+        assert!(consume("cli-test-1", &issued.token));
+    }
+
+    #[test]
+    fn issued_token_rejects_wrong_client_id() {
+        let issued = issue("cli-test-2", None).expect("issue");
+        assert!(!consume("attacker-id", &issued.token));
+    }
+
+    #[test]
+    fn wrong_client_id_does_not_consume_token() {
+        // Mismatched consume must leave the token intact so the legitimate
+        // subscriber can still validate after the failed probe — otherwise
+        // a wrong-id request becomes a one-shot DoS.
+        let issued = issue("cli-test-mismatch", None).expect("issue");
+        assert!(!consume("attacker-id", &issued.token));
+        assert!(
+            consume("cli-test-mismatch", &issued.token),
+            "legitimate consume must still succeed after a mismatched probe"
+        );
+    }
+
+    #[test]
+    fn consumed_token_cannot_be_reused() {
+        let issued = issue("cli-test-3", None).expect("issue");
+        assert!(consume("cli-test-3", &issued.token));
+        assert!(
+            !consume("cli-test-3", &issued.token),
+            "tokens must be single-shot"
+        );
+    }
+
+    #[test]
+    fn expired_token_is_rejected() {
+        let issued = issue("cli-test-4", Some(Duration::from_millis(1))).expect("issue");
+        std::thread::sleep(Duration::from_millis(20));
+        assert!(!consume("cli-test-4", &issued.token));
+    }
+
+    #[test]
+    fn unknown_token_is_rejected() {
+        assert!(!consume("any-id", "f00ba1"));
+    }
+
+    #[test]
+    fn ttl_override_is_clamped_to_max() {
+        // Any caller asking for more than `MAX_TTL` collapses to the cap;
+        // confirm the issue path does not panic and the resulting token
+        // still validates.
+        let issued =
+            issue("cli-test-clamp", Some(Duration::from_secs(60 * 60 * 24))).expect("issue");
+        assert!(issued.valid_until <= Instant::now() + MAX_TTL + Duration::from_secs(1));
+        assert!(consume("cli-test-clamp", &issued.token));
+    }
+}
diff --git a/src/core/event_bus/events.rs b/src/core/event_bus/events.rs
index 52b0870ad..948acd92c 100644
--- a/src/core/event_bus/events.rs
+++ b/src/core/event_bus/events.rs
@@ -92,10 +92,22 @@ pub enum DomainEvent {
 
     // ── Channels ────────────────────────────────────────────────────────
     /// An inbound channel message from the transport layer, ready for processing.
+    ///
+    /// `sender`, `reply_target`, and `thread_ts` are carried alongside
+    /// `channel` so the agent loop can derive per-sender conversation keys
+    /// the same way `channels::context::conversation_history_key` does for
+    /// other inbound paths — keying on `channel` alone collapses distinct
+    /// senders inside a shared channel into one cached session.
     ChannelInboundMessage {
         event_name: String,
         channel: String,
         message: String,
+        #[doc = "Originating user/account id within the channel. `None` for legacy publishers that don't surface it."]
+        sender: Option<String>,
+        #[doc = "Direct-message peer or group thread the reply should go to. `None` when the channel does not distinguish."]
+        reply_target: Option<String>,
+        #[doc = "Slack/Discord thread anchor when the message is in-thread. `None` for top-level messages."]
+        thread_ts: Option<String>,
         raw_data: serde_json::Value,
     },
     /// A message was received on a channel.
diff --git a/src/core/event_bus/events_tests.rs b/src/core/event_bus/events_tests.rs
index acdd05170..424db45fb 100644
--- a/src/core/event_bus/events_tests.rs
+++ b/src/core/event_bus/events_tests.rs
@@ -79,6 +79,9 @@ fn all_variants_have_correct_domain() {
                 event_name: "telegram:message".into(),
                 channel: "telegram".into(),
                 message: "hi".into(),
+                sender: None,
+                reply_target: None,
+                thread_ts: None,
                 raw_data: serde_json::Value::Null,
             },
             "channel",
diff --git a/src/core/jsonrpc.rs b/src/core/jsonrpc.rs
index 1cfb93a74..a80e6a2d2 100644
--- a/src/core/jsonrpc.rs
+++ b/src/core/jsonrpc.rs
@@ -460,11 +460,95 @@ fn error_html(message: &str) -> String {
     )
 }
 
+/// Inspect the browser fetch-metadata + Referer/Origin headers and decide
+/// whether the inbound `/auth/telegram` request looks like a legitimate
+/// top-level redirect from Telegram, or a cross-site CSRF attempt.
+///
+/// The endpoint cannot require a bearer token (the redirect happens in a
+/// fresh browser tab; `EventSource`-style header injection is not an
+/// option), and there is no in-process state issued by an authenticated
+/// FE flow today (`/start register` is initiated in Telegram, not in the
+/// local app). So this fetch-metadata gate is the layer that distinguishes
+/// "user clicked the link the bot sent them" from "malicious page
+/// navigates the user's loopback core via `window.location`/`<img>`".
+///
+/// Accepted shapes:
+/// - All `Sec-Fetch-*` headers absent (older browsers, CLI clients).
+/// - `Sec-Fetch-Mode: navigate` AND `Sec-Fetch-Dest: document`.
+/// - `Sec-Fetch-Site` is `same-origin` / `none`, OR `cross-site` with a
+///   `Referer` that starts with `https://t.me/` (the legit bot redirect).
+///
+/// Rejected shapes:
+/// - `Sec-Fetch-Mode` is `no-cors` / `cors` / `same-origin` (only
+///   `navigate` makes sense for a top-level page load).
+/// - `Sec-Fetch-Dest` is anything other than `document` (image/script/
+///   iframe embeds from malicious pages).
+/// - `Sec-Fetch-Site: cross-site` with a `Referer`/`Origin` that is not
+///   `https://t.me/...` (CSRF redirect from a third-party site).
+fn telegram_callback_origin_ok(headers: &axum::http::HeaderMap) -> Result<(), &'static str> {
+    let get_str = |name: &str| -> Option<&str> {
+        headers
+            .get(name)
+            .and_then(|v| v.to_str().ok())
+            .map(|s| s.trim())
+            .filter(|s| !s.is_empty())
+    };
+
+    let mode = get_str("sec-fetch-mode");
+    let dest = get_str("sec-fetch-dest");
+    let site = get_str("sec-fetch-site");
+    let referer = get_str("referer");
+    let origin = get_str("origin");
+
+    if let Some(mode) = mode {
+        if mode != "navigate" {
+            return Err("Sec-Fetch-Mode must be 'navigate'");
+        }
+    }
+    if let Some(dest) = dest {
+        if dest != "document" {
+            return Err("Sec-Fetch-Dest must be 'document'");
+        }
+    }
+
+    let referer_is_telegram = referer
+        .map(|r| r.starts_with("https://t.me/") || r.starts_with("https://web.telegram.org/"))
+        .unwrap_or(false);
+    let origin_is_telegram = origin
+        .map(|o| o == "https://t.me" || o == "https://web.telegram.org")
+        .unwrap_or(false);
+
+    if let Some(site) = site {
+        if site == "cross-site" && !(referer_is_telegram || origin_is_telegram) {
+            return Err("cross-site redirect must originate from telegram");
+        }
+    } else if let Some(referer) = referer {
+        // No Sec-Fetch-Site: fall back to Referer host check. Accept
+        // loopback referer (direct nav inside the local app) — parsed
+        // exactly so `http://localhost.attacker.example/...` does not
+        // satisfy the gate — and accept telegram referer (legit bot
+        // redirect); reject everything else.
+        let local = url::Url::parse(referer)
+            .ok()
+            .and_then(|u| u.host_str().map(str::to_string))
+            .map(|h| matches!(h.as_str(), "localhost" | "127.0.0.1" | "::1"))
+            .unwrap_or(false);
+        if !(local || referer_is_telegram) {
+            return Err("Referer must be telegram or local");
+        }
+    }
+
+    Ok(())
+}
+
 /// Handles the Telegram authentication callback.
 ///
 /// It consumes a one-time token, exchanges it for a JWT from the backend,
 /// and stores the session locally.
-async fn telegram_auth_handler(Query(query): Query<TelegramAuthQuery>) -> impl IntoResponse {
+async fn telegram_auth_handler(
+    headers: axum::http::HeaderMap,
+    Query(query): Query<TelegramAuthQuery>,
+) -> impl IntoResponse {
     let html_response = |status: StatusCode, body: String| -> Response {
         (
             status,
@@ -474,6 +558,18 @@ async fn telegram_auth_handler(Query(query): Query<TelegramAuthQuery>) -> impl I
             .into_response()
     };
 
+    if let Err(reason) = telegram_callback_origin_ok(&headers) {
+        log::warn!("[auth:telegram] rejecting callback: {reason}");
+        return html_response(
+            StatusCode::FORBIDDEN,
+            error_html(
+                "This login callback did not come from the Telegram bot. \
+                 Open the link the bot sent you directly, do not let \
+                 another page redirect you here.",
+            ),
+        );
+    }
+
     let token = match query
         .token
         .as_deref()
@@ -802,36 +898,107 @@ async fn schema_handler(State(_state): State<AppState>) -> impl IntoResponse {
 }
 
 /// Query parameters for the events SSE endpoint.
+///
+/// `client_id` selects which broadcast events to forward; `token` is the
+/// single-shot bind token minted by the `core.events_subscribe_token` RPC.
+/// Both are required — browser `EventSource` cannot attach an
+/// `Authorization` header, so the bind token is the only credential the
+/// endpoint accepts.
 #[derive(Debug, serde::Deserialize)]
 struct EventsQuery {
-    /// Unique identifier for the client requesting events.
     client_id: String,
+    #[serde(default)]
+    token: Option<String>,
 }
 
 /// Handler for the main events SSE endpoint.
 ///
-/// Streams real-time events filtered by `client_id`.
+/// Accepts either of two credentials:
+/// 1. `Authorization: Bearer <core token>` — used by CLI tooling, the
+///    Tauri shell via `core_rpc_relay`, and the in-tree e2e suite that
+///    can set HTTP headers directly. Validated against the same
+///    per-process bearer the rest of `/rpc` uses.
+/// 2. `?token=<bind>` minted via the `core.events_subscribe_token` RPC
+///    — used by browser `EventSource`, which cannot attach custom
+///    headers. The token is bound to a specific `client_id` and is
+///    consumed on validation so a leaked URL cannot be replayed.
+///
+/// Both paths converge on the same broadcast stream filtered by
+/// `client_id`.
 async fn events_handler(
+    headers: axum::http::HeaderMap,
     Query(query): Query<EventsQuery>,
-) -> Sse<impl tokio_stream::Stream<Item = Result<Event, std::convert::Infallible>>> {
+) -> Response {
+    let bearer = headers
+        .get(header::AUTHORIZATION)
+        .and_then(|v| v.to_str().ok())
+        .and_then(|v| v.strip_prefix("Bearer "))
+        .map(str::trim)
+        .filter(|s| !s.is_empty());
+    let bearer_ok = bearer
+        .map(crate::core::auth::verify_bearer_token)
+        .unwrap_or(false);
+
+    if !bearer_ok {
+        let supplied_token = query
+            .token
+            .as_deref()
+            .map(str::trim)
+            .filter(|s| !s.is_empty());
+        let Some(supplied_token) = supplied_token else {
+            log::warn!(
+                "[events] reject subscribe: missing bind token + missing bearer (client_id_len={})",
+                query.client_id.len()
+            );
+            return (
+                StatusCode::UNAUTHORIZED,
+                Json(json!({
+                    "ok": false,
+                    "error": "unauthorized",
+                    "message": "Missing credentials. Supply 'Authorization: Bearer <core>' or mint a bind token with the `core.events_subscribe_token` RPC and pass it as ?token="
+                })),
+            )
+                .into_response();
+        };
+        if !crate::core::event_bind_tokens::consume(&query.client_id, supplied_token) {
+            log::warn!(
+                "[events] reject subscribe: bind token invalid or expired (client_id_len={})",
+                query.client_id.len()
+            );
+            return (
+                StatusCode::UNAUTHORIZED,
+                Json(json!({
+                    "ok": false,
+                    "error": "unauthorized",
+                    "message": "Bind token is unknown, expired, or bound to a different client_id."
+                })),
+            )
+                .into_response();
+        }
+    }
+
     let client_id = query.client_id;
     let rx = crate::openhuman::channels::providers::web::subscribe_web_channel_events();
-    let stream = tokio_stream::wrappers::BroadcastStream::new(rx).filter_map(move |item| {
-        let event = match item {
-            Ok(ev) => ev,
-            Err(_) => return None,
-        };
-        if event.client_id != client_id {
-            return None;
-        }
-        let data = match serde_json::to_string(&event) {
-            Ok(data) => data,
-            Err(_) => return None,
-        };
-        Some(Ok(Event::default().event(event.event).data(data)))
-    });
+    let stream = tokio_stream::wrappers::BroadcastStream::new(rx).filter_map(
+        move |item| -> Option<Result<Event, std::convert::Infallible>> {
+            let event = match item {
+                Ok(ev) => ev,
+                Err(_) => return None,
+            };
+            if event.client_id != client_id {
+                return None;
+            }
+            let data = match serde_json::to_string(&event) {
+                Ok(data) => data,
+                Err(_) => return None,
+            };
+            Some(Ok(Event::default().event(event.event).data(data)))
+        },
+    );
 
-    Sse::new(stream).keep_alive(KeepAlive::new().interval(std::time::Duration::from_secs(10)))
+    Sse::new(stream)
+        .keep_alive(KeepAlive::new().interval(std::time::Duration::from_secs(10)))
+        .into_response()
 }
 
 /// Handler for the webhook debug events SSE endpoint.
@@ -862,7 +1029,7 @@ async fn root_handler() -> impl IntoResponse {
             "endpoints": {
                 "health": "/health",
                 "schema": "/schema",
-                "events": "/events?client_id=<id>",
+                "events": "/events?client_id=<id>&token=<core.events_subscribe_token>",
                 "rpc": "/rpc"
             },
             "usage": {
diff --git a/src/core/jsonrpc_tests.rs b/src/core/jsonrpc_tests.rs
index d404d83fc..6e91b9225 100644
--- a/src/core/jsonrpc_tests.rs
+++ b/src/core/jsonrpc_tests.rs
@@ -933,6 +933,98 @@ fn escape_html_is_noop_for_safe_text() {
     assert_eq!(escape_html(""), "");
 }
 
+// --- telegram callback fetch-metadata gate --------------------------------
+
+fn hdr_map(pairs: &[(&str, &str)]) -> axum::http::HeaderMap {
+    let mut m = axum::http::HeaderMap::new();
+    for (k, v) in pairs {
+        m.insert(
+            axum::http::HeaderName::from_bytes(k.as_bytes()).unwrap(),
+            axum::http::HeaderValue::from_str(v).unwrap(),
+        );
+    }
+    m
+}
+
+#[test]
+fn telegram_callback_origin_ok_accepts_no_metadata_headers() {
+    // Older browsers and CLI clients (curl) send neither Sec-Fetch-* nor
+    // Origin/Referer. The legacy flow has to keep working — reject only
+    // when there is evidence of a cross-site embedded context.
+    let headers = hdr_map(&[]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_ok());
+}
+
+#[test]
+fn telegram_callback_origin_ok_accepts_legit_top_nav_from_telegram() {
+    let headers = hdr_map(&[
+        ("sec-fetch-mode", "navigate"),
+        ("sec-fetch-dest", "document"),
+        ("sec-fetch-site", "cross-site"),
+        ("referer", "https://t.me/some_bot"),
+    ]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_ok());
+}
+
+#[test]
+fn telegram_callback_origin_ok_accepts_same_origin_local_nav() {
+    let headers = hdr_map(&[
+        ("sec-fetch-mode", "navigate"),
+        ("sec-fetch-dest", "document"),
+        ("sec-fetch-site", "same-origin"),
+    ]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_ok());
+}
+
+#[test]
+fn telegram_callback_origin_ok_rejects_image_embed() {
+    let headers = hdr_map(&[
+        ("sec-fetch-mode", "no-cors"),
+        ("sec-fetch-dest", "image"),
+        ("sec-fetch-site", "cross-site"),
+    ]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+}
+
+#[test]
+fn telegram_callback_origin_ok_rejects_iframe_embed() {
+    let headers = hdr_map(&[
+        ("sec-fetch-mode", "navigate"),
+        ("sec-fetch-dest", "iframe"),
+        ("sec-fetch-site", "cross-site"),
+    ]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+}
+
+#[test]
+fn telegram_callback_origin_ok_rejects_cross_site_from_non_telegram() {
+    let headers = hdr_map(&[
+        ("sec-fetch-mode", "navigate"),
+        ("sec-fetch-dest", "document"),
+        ("sec-fetch-site", "cross-site"),
+        ("referer", "https://attacker.example/page"),
+    ]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+}
+
+#[test]
+fn telegram_callback_origin_ok_rejects_non_telegram_referer_without_fetch_metadata() {
+    let headers = hdr_map(&[("referer", "https://attacker.example/post")]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+}
+
+#[test]
+fn telegram_callback_origin_ok_rejects_localhost_host_prefix_decoy() {
+    // Regression: prefix-matching the referer accepted hostnames like
+    // `http://localhost.attacker.example/...`. With exact-host parsing
+    // these must be rejected even when no fetch-metadata headers are
+    // present.
+    let headers = hdr_map(&[("referer", "http://localhost.attacker.example/cb")]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+    let headers = hdr_map(&[("referer", "http://127.0.0.1.attacker.example/cb")]);
+    assert!(super::telegram_callback_origin_ok(&headers).is_err());
+}
+
 // --- invoke_method parameter-shape errors ---------------------------------
 
 #[tokio::test]
diff --git a/src/core/mod.rs b/src/core/mod.rs
index 1d93d0cfe..a7ae071c7 100644
--- a/src/core/mod.rs
+++ b/src/core/mod.rs
@@ -12,6 +12,7 @@ pub mod auth;
 pub mod autocomplete_cli_adapter;
 pub mod cli;
 pub mod dispatch;
+pub mod event_bind_tokens;
 pub mod event_bus;
 pub mod jsonrpc;
 pub mod legacy_aliases;
diff --git a/src/core/socketio.rs b/src/core/socketio.rs
index 3e723441b..fd91122af 100644
--- a/src/core/socketio.rs
+++ b/src/core/socketio.rs
@@ -1,9 +1,81 @@
 use serde::Deserialize;
 use serde::Serialize;
 use serde_json::json;
-use socketioxide::extract::{Data, SocketRef};
+use socketioxide::extract::{Data, SocketRef, TryData};
 use socketioxide::SocketIo;
 
+/// Marker stored in [`SocketRef::extensions`] once a connection has presented a
+/// bearer token that matches the active per-process RPC token.
+///
+/// Event handlers consult this before forwarding attacker-controllable input
+/// into the JSON-RPC dispatcher or the web-chat orchestrator: an unauthenticated
+/// socket that never picked up the marker is allowed to receive broadcast-style
+/// events (read-only) but cannot trigger executable work.
+#[derive(Clone, Copy, Debug)]
+struct AuthedConnection;
+
+/// Connection-time payload the client passes via Socket.IO's `auth` field.
+///
+/// Browsers do not let `EventSource` / `WebSocket` clients attach custom
+/// headers, so the handshake `auth` map is the only header-equivalent slot
+/// available for our per-process bearer. The socket-IO Node/JS clients all
+/// surface `io(url, { auth: { token: "<hex>" } })` for this.
+#[derive(Debug, Default, Deserialize)]
+struct HandshakeAuth {
+    #[serde(default)]
+    token: Option<String>,
+}
+
+/// Origins the local core trusts at the Socket.IO handshake.
+///
+/// `tauri://localhost` is the production app webview; `http://localhost:*`
+/// and `http://127.0.0.1:*` cover the Vite dev server (`pnpm dev:app`)
+/// and standalone CLI tooling that opens browser pages against the local
+/// listener. A missing `Origin` header is treated as a native (non-browser)
+/// client and accepted — only the cross-origin browser-page case is the
+/// targeted bad actor here.
+fn origin_is_allowed(origin: Option<&str>) -> bool {
+    let Some(origin) = origin else {
+        return true; // native clients (CLI, Tauri shell) — no Origin header
+    };
+    let origin = origin.trim();
+    if origin.is_empty() || origin == "null" {
+        return false;
+    }
+    if origin == "tauri://localhost" || origin == "https://tauri.localhost" {
+        return true;
+    }
+    // Parse the URL and compare the host EXACTLY against the loopback
+    // allowlist — `starts_with` matching accepted decoys like
+    // `http://localhost.attacker.example` and bypassed the gate.
+    let Ok(parsed) = url::Url::parse(origin) else {
+        return false;
+    };
+    // `url::Url::host_str` returns IPv6 hosts with surrounding brackets,
+    // hostnames bare. Accept both shapes.
+    matches!(
+        parsed.host_str(),
+        Some("localhost" | "127.0.0.1" | "::1" | "[::1]")
+    )
+}
+
+/// True when `socket` finished the handshake with a valid bearer token.
+fn socket_is_authed(socket: &SocketRef) -> bool {
+    socket.extensions.get::<AuthedConnection>().is_some()
+}
+
+/// Best-effort disconnect. Called when we discover an unauthenticated socket
+/// inside an event handler — the connect path already disconnects the bad
+/// origins / wrong tokens, so this is purely a defense-in-depth path.
+fn drop_unauthed(socket: &SocketRef, reason: &'static str) {
+    log::warn!(
+        "[socketio] dropping unauthenticated socket id={} reason={}",
+        socket.id,
+        reason
+    );
+    let _ = socket.clone().disconnect();
+}
+
 /// Standard event payload for the web channel transport.
 ///
 /// This structure defines the data sent to Socket.IO clients for various
@@ -181,129 +253,179 @@ pub fn attach_socketio() -> (socketioxide::layer::SocketIoLayer, SocketIo) {
         io.config().engine_config.req_path
     );
 
-    io.ns("/", |socket: SocketRef| {
-        let client_id = socket.id.to_string();
-        log::info!("[socketio] client connected id={client_id}");
-        // Join a room named after the client ID for targeted event delivery.
-        join_room_logged(&socket, &client_id, &client_id);
-        // Also auto-join the "system" room so every connected client
-        // receives broadcast-style events that aren't tied to a
-        // specific chat thread. Today this covers proactive messages
-        // (welcome agent, morning briefing, cron-driven announcements)
-        // which `channels::proactive::ProactiveMessageSubscriber`
-        // emits with `client_id = "system"` — see `emit_web_channel_event`.
-        // If this join fails the welcome message silently disappears,
-        // so we log both success and failure for diagnosability.
-        join_room_logged(&socket, "system", &client_id);
-        let ready_payload = json!({ "sid": client_id });
-        log::debug!("[socketio] emit event=ready to_client={}", socket.id);
-        let _ = socket.emit("ready", &ready_payload);
+    io.ns(
+        "/",
+        |socket: SocketRef, TryData(handshake): TryData<HandshakeAuth>| {
+            let client_id = socket.id.to_string();
 
-        // Handler for JSON-RPC over WebSocket.
-        socket.on(
-            "rpc:request",
-            |socket: SocketRef, Data(payload): Data<SocketRpcRequest>| async move {
-                let client_id = socket.id.to_string();
-                log::info!(
-                    "[socketio] rpc:request method={} id={} client={}",
-                    payload.method,
-                    payload.id,
+            // Reject cross-origin browser pages before the handshake completes.
+            // Native clients (Tauri shell, CLI) do not set an `Origin` header and
+            // are accepted; only browser pages from origins outside the local
+            // app surface are dropped here. See `origin_is_allowed`.
+            let origin = socket
+                .req_parts()
+                .headers
+                .get(axum::http::header::ORIGIN)
+                .and_then(|v| v.to_str().ok())
+                .map(|s| s.to_string());
+            if !origin_is_allowed(origin.as_deref()) {
+                log::warn!(
+                    "[socketio] rejecting connect: bad origin {:?} client={}",
+                    origin,
                     client_id
                 );
+                let _ = socket.clone().disconnect();
+                return;
+            }
 
-                // Invoke the method through the same logic used by the HTTP RPC endpoint.
-                let response = match crate::core::jsonrpc::invoke_method(
-                    crate::core::jsonrpc::default_state(),
-                    payload.method.as_str(),
-                    payload.params,
-                )
-                .await
-                {
-                    Ok(result) => (
-                        "rpc:response",
-                        json!({ "id": payload.id, "result": result }),
-                    ),
-                    Err(message) => (
-                        "rpc:error",
-                        json!({
-                            "id": payload.id,
-                            "error": { "code": -32000, "message": message }
-                        }),
-                    ),
-                };
+            // Verify the handshake bearer matches the per-process RPC token.
+            // `TryData` lets us treat a missing/malformed `auth` payload as a
+            // soft failure (no panic) and reject the connect cleanly.
+            let supplied = handshake.ok().and_then(|h| h.token).unwrap_or_default();
+            if !crate::core::auth::verify_bearer_token(&supplied) {
+                log::warn!(
+                    "[socketio] rejecting connect: missing or invalid bearer client={}",
+                    client_id
+                );
+                let _ = socket.clone().disconnect();
+                return;
+            }
+            socket.extensions.insert(AuthedConnection);
 
-                let _ = socket.emit(response.0, &response.1);
-            },
-        );
+            log::info!("[socketio] client connected id={client_id} (authenticated)");
+            // Join a room named after the client ID for targeted event delivery.
+            join_room_logged(&socket, &client_id, &client_id);
+            // Also auto-join the "system" room so every connected client
+            // receives broadcast-style events that aren't tied to a
+            // specific chat thread. Today this covers proactive messages
+            // (welcome agent, morning briefing, cron-driven announcements)
+            // which `channels::proactive::ProactiveMessageSubscriber`
+            // emits with `client_id = "system"` — see `emit_web_channel_event`.
+            // If this join fails the welcome message silently disappears,
+            // so we log both success and failure for diagnosability.
+            join_room_logged(&socket, "system", &client_id);
+            let ready_payload = json!({ "sid": client_id });
+            log::debug!("[socketio] emit event=ready to_client={}", socket.id);
+            let _ = socket.emit("ready", &ready_payload);
 
-        // Handler for starting a chat turn.
-        socket.on(
-            "chat:start",
-            |socket: SocketRef, Data(payload): Data<ChatStartPayload>| async move {
-                let client_id = socket.id.to_string();
-                let thread_id = payload.thread_id.clone();
-                let model_override = payload.model_override.or(payload.model);
-                log::debug!(
+            // Handler for JSON-RPC over WebSocket.
+            socket.on(
+                "rpc:request",
+                |socket: SocketRef, Data(payload): Data<SocketRpcRequest>| async move {
+                    if !socket_is_authed(&socket) {
+                        drop_unauthed(&socket, "rpc:request from unauthenticated socket");
+                        return;
+                    }
+                    let client_id = socket.id.to_string();
+                    log::info!(
+                        "[socketio] rpc:request method={} id={} client={}",
+                        payload.method,
+                        payload.id,
+                        client_id
+                    );
+
+                    // Invoke the method through the same logic used by the HTTP RPC endpoint.
+                    let response = match crate::core::jsonrpc::invoke_method(
+                        crate::core::jsonrpc::default_state(),
+                        payload.method.as_str(),
+                        payload.params,
+                    )
+                    .await
+                    {
+                        Ok(result) => (
+                            "rpc:response",
+                            json!({ "id": payload.id, "result": result }),
+                        ),
+                        Err(message) => (
+                            "rpc:error",
+                            json!({
+                                "id": payload.id,
+                                "error": { "code": -32000, "message": message }
+                            }),
+                        ),
+                    };
+
+                    let _ = socket.emit(response.0, &response.1);
+                },
+            );
+
+            // Handler for starting a chat turn.
+            socket.on(
+                "chat:start",
+                |socket: SocketRef, Data(payload): Data<ChatStartPayload>| async move {
+                    if !socket_is_authed(&socket) {
+                        drop_unauthed(&socket, "chat:start from unauthenticated socket");
+                        return;
+                    }
+                    let client_id = socket.id.to_string();
+                    let thread_id = payload.thread_id.clone();
+                    let model_override = payload.model_override.or(payload.model);
+                    log::debug!(
                     "[socketio] recv event=chat:start client_id={} thread_id={} message_bytes={}",
                     client_id,
                     thread_id,
                     payload.message.len()
                 );
 
-                // Trigger the web channel's chat logic.
-                match crate::openhuman::channels::providers::web::start_chat(
-                    &client_id,
-                    &payload.thread_id,
-                    &payload.message,
-                    model_override,
-                    payload.temperature,
-                    payload.profile_id,
-                    payload.locale,
-                )
-                .await
-                {
-                    Ok(request_id) => {
-                        let accepted_payload = json!({
-                            "event": "chat_accepted",
-                            "client_id": client_id,
-                            "thread_id": thread_id,
-                            "request_id": request_id,
-                        });
-                        emit_with_aliases(&socket, "chat_accepted", &accepted_payload);
+                    // Trigger the web channel's chat logic.
+                    match crate::openhuman::channels::providers::web::start_chat(
+                        &client_id,
+                        &payload.thread_id,
+                        &payload.message,
+                        model_override,
+                        payload.temperature,
+                        payload.profile_id,
+                        payload.locale,
+                    )
+                    .await
+                    {
+                        Ok(request_id) => {
+                            let accepted_payload = json!({
+                                "event": "chat_accepted",
+                                "client_id": client_id,
+                                "thread_id": thread_id,
+                                "request_id": request_id,
+                            });
+                            emit_with_aliases(&socket, "chat_accepted", &accepted_payload);
+                        }
+                        Err(error) => {
+                            let error_payload = json!({
+                                "event": "chat_error",
+                                "client_id": client_id,
+                                "thread_id": thread_id,
+                                "request_id": "",
+                                "message": error,
+                                "error_type": "inference",
+                            });
+                            emit_with_aliases(&socket, "chat_error", &error_payload);
+                        }
                     }
-                    Err(error) => {
-                        let error_payload = json!({
-                            "event": "chat_error",
-                            "client_id": client_id,
-                            "thread_id": thread_id,
-                            "request_id": "",
-                            "message": error,
-                            "error_type": "inference",
-                        });
-                        emit_with_aliases(&socket, "chat_error", &error_payload);
-                    }
-                }
-            },
-        );
+                },
+            );
 
-        // Handler for cancelling an active chat turn.
-        socket.on(
-            "chat:cancel",
-            |socket: SocketRef, Data(payload): Data<ChatCancelPayload>| async move {
-                let client_id = socket.id.to_string();
-                log::debug!(
-                    "[socketio] recv event=chat:cancel client_id={} thread_id={}",
-                    client_id,
-                    payload.thread_id
-                );
-                let _ = crate::openhuman::channels::providers::web::cancel_chat(
-                    &client_id,
-                    &payload.thread_id,
-                )
-                .await;
-            },
-        );
-    });
+            // Handler for cancelling an active chat turn.
+            socket.on(
+                "chat:cancel",
+                |socket: SocketRef, Data(payload): Data<ChatCancelPayload>| async move {
+                    if !socket_is_authed(&socket) {
+                        drop_unauthed(&socket, "chat:cancel from unauthenticated socket");
+                        return;
+                    }
+                    let client_id = socket.id.to_string();
+                    log::debug!(
+                        "[socketio] recv event=chat:cancel client_id={} thread_id={}",
+                        client_id,
+                        payload.thread_id
+                    );
+                    let _ = crate::openhuman::channels::providers::web::cancel_chat(
+                        &client_id,
+                        &payload.thread_id,
+                    )
+                    .await;
+                },
+            );
+        },
+    );
 
     (layer, io)
 }
@@ -615,7 +737,7 @@ fn emit_room_with_aliases(io: &SocketIo, room: &str, name: &str, payload: &serde
 
 #[cfg(test)]
 mod tests {
-    use super::event_alias;
+    use super::{event_alias, origin_is_allowed};
 
     #[test]
     fn event_alias_translates_between_delimiters() {
@@ -623,4 +745,49 @@ mod tests {
         assert_eq!(event_alias("chat:error").as_deref(), Some("chat_error"));
         assert_eq!(event_alias("ready"), None);
     }
+
+    #[test]
+    fn origin_allowlist_accepts_native_clients() {
+        assert!(origin_is_allowed(None));
+    }
+
+    #[test]
+    fn origin_allowlist_accepts_tauri_localhost() {
+        assert!(origin_is_allowed(Some("tauri://localhost")));
+        assert!(origin_is_allowed(Some("https://tauri.localhost")));
+    }
+
+    #[test]
+    fn origin_allowlist_accepts_local_dev_server() {
+        assert!(origin_is_allowed(Some("http://localhost:1420")));
+        assert!(origin_is_allowed(Some("http://127.0.0.1:1420")));
+        assert!(origin_is_allowed(Some("http://[::1]:1420")));
+    }
+
+    #[test]
+    fn origin_allowlist_rejects_cross_origin_browser_pages() {
+        assert!(!origin_is_allowed(Some("https://attacker.example")));
+        assert!(!origin_is_allowed(Some("http://evil.local")));
+        assert!(!origin_is_allowed(Some("null")));
+        assert!(!origin_is_allowed(Some("")));
+    }
+
+    #[test]
+    fn origin_allowlist_rejects_host_prefix_decoys() {
+        // Regression: `starts_with("localhost")` accepted these; the exact
+        // host match must not.
+        assert!(!origin_is_allowed(Some(
+            "http://localhost.attacker.example"
+        )));
+        assert!(!origin_is_allowed(Some(
+            "http://127.0.0.1.attacker.example"
+        )));
+        assert!(!origin_is_allowed(Some("https://localhost-evil")));
+    }
+
+    #[test]
+    fn origin_allowlist_rejects_unparseable_origin() {
+        assert!(!origin_is_allowed(Some("not a url")));
+        assert!(!origin_is_allowed(Some("javascript:alert(1)")));
+    }
 }
diff --git a/src/openhuman/agent/harness/memory_context.rs b/src/openhuman/agent/harness/memory_context.rs
index 620734209..a5ca26e25 100644
--- a/src/openhuman/agent/harness/memory_context.rs
+++ b/src/openhuman/agent/harness/memory_context.rs
@@ -1,3 +1,4 @@
+use super::memory_context_safety::{is_potentially_untrusted, wrap_untrusted_for_agent};
 use crate::openhuman::memory::Memory;
 use crate::openhuman::util::provenance_tag;
 use std::collections::HashSet;
@@ -58,7 +59,13 @@ pub(crate) async fn build_context(
             context.push_str("[Memory context]\n");
             for entry in &relevant {
                 seen_keys.insert(entry.key.clone());
-                let _ = writeln!(context, "- {}: {}", entry.key, entry.content);
+                let rendered_content = if is_potentially_untrusted(entry) {
+                    let hint = entry.namespace.as_deref().unwrap_or("connector");
+                    wrap_untrusted_for_agent(&entry.content, hint)
+                } else {
+                    entry.content.clone()
+                };
+                let _ = writeln!(context, "- {}: {}", entry.key, rendered_content);
             }
             context.push('\n');
         }
diff --git a/src/openhuman/agent/harness/memory_context_safety.rs b/src/openhuman/agent/harness/memory_context_safety.rs
new file mode 100644
index 000000000..46e61a720
--- /dev/null
+++ b/src/openhuman/agent/harness/memory_context_safety.rs
@@ -0,0 +1,251 @@
+//! Trust-tier helpers for memory entries surfaced into agent prompts.
+//!
+//! Memory entries reach the agent prompt by way of vector-recall over the
+//! full memory store, which mixes content from many provenance tiers:
+//!
+//! - **User-authored** turns from the same chat (high trust).
+//! - **Agent-authored** summaries and working-memory snapshots (high trust).
+//! - **Connector-synced** content harvested from Gmail / Slack / Notion /
+//!   Discord / web feeds (untrusted: anything in the body of an email, the
+//!   text of a Slack DM, or a Notion page is text the agent has no a-priori
+//!   reason to obey).
+//!
+//! Recall returns the same shape regardless of which tier the row came
+//! from, so a prompt-injection paragraph that lives inside an inbound
+//! email reaches the agent's working context with the same visual weight
+//! as a system-issued instruction. This module is the narrowest possible
+//! mitigation: a heuristic that flags potentially-untrusted entries by
+//! namespace / key shape, and a wrapping helper that surrounds the entry
+//! with explicit `<untrusted-source>` markers so the safety preamble and
+//! the model itself have a fighting chance of distinguishing context from
+//! instructions.
+//!
+//! A proper fix is a typed `Provenance` enum carried on every memory row,
+//! populated by the ingestion pipeline. That requires a schema migration
+//! across `MemoryEntry`, the SQLite store, and every namespace creator —
+//! out of scope for this commit. The heuristics here intentionally err
+//! toward over-wrapping: it is safer to tag a user-authored row as
+//! untrusted than to leave a connector-synced one bare.
+
+use crate::openhuman::memory::MemoryEntry;
+
+/// Conservative classifier — returns `true` when the entry is unlikely to
+/// be locally-authored and therefore SHOULD be wrapped before reaching
+/// the agent prompt.
+///
+/// Rules (any match flips to untrusted):
+/// - Namespace exists and is not one of the local-authored short-list
+///   (`working`, `agent`, `local`, `core`, `global`, `default`, or the
+///   ingestion-internal `tree.*` namespaces that are summarised locally).
+/// - Key carries a known connector prefix (`chat:`, `email:`, `notion:`,
+///   `drive:`, `discord:`, `telegram:`, `whatsapp:`, `slack:`, `gmail:`,
+///   `outlook:`, `imap:`, `meeting:`, `web:`).
+///
+/// Local-authored namespaces are an allowlist so an unrecognised namespace
+/// surfaces as "untrusted" (default-deny). The mitigation is conservative
+/// on purpose; refining it requires explicit provenance tagging at
+/// ingest time.
+pub fn is_potentially_untrusted(entry: &MemoryEntry) -> bool {
+    if let Some(ns) = entry.namespace.as_deref() {
+        let ns = ns.trim().to_ascii_lowercase();
+        if !is_locally_authored_namespace(&ns) {
+            return true;
+        }
+    }
+
+    let key_lower = entry.key.to_ascii_lowercase();
+    let connector_prefixes: &[&str] = &[
+        "chat:",
+        "email:",
+        "notion:",
+        "drive:",
+        "discord:",
+        "telegram:",
+        "whatsapp:",
+        "slack:",
+        "gmail:",
+        "outlook:",
+        "imap:",
+        "meeting:",
+        "web:",
+    ];
+    connector_prefixes.iter().any(|p| key_lower.starts_with(p))
+}
+
+fn is_locally_authored_namespace(ns: &str) -> bool {
+    // Exact-match short list — everything else (including ingestion-derived
+    // namespaces) is treated as untrusted by default.
+    matches!(
+        ns,
+        "working" | "agent" | "local" | "core" | "global" | "default" | "user"
+    ) || ns.starts_with("working.")
+        || ns.starts_with("agent.")
+        || ns.starts_with("tree.")
+}
+
+/// Wrap `content` in explicit untrusted-source markers so the agent
+/// prompt visually distinguishes it from system instructions.
+///
+/// `source_hint` is a short, human-readable hint (`"gmail"`, `"slack"`,
+/// `"connector"`, `"recall"`, …) that lands in the tag attributes so the
+/// model can see which surface produced the row without revealing
+/// content that should not leave the trust boundary.
+///
+/// Both `source_hint` and `content` are sanitised before they reach the
+/// formatted string — without sanitisation a payload containing a
+/// literal `</untrusted-source>` or stray quote could close or forge
+/// the marker and slip back into the trusted region.
+pub fn wrap_untrusted_for_agent(content: &str, source_hint: &str) -> String {
+    let hint = sanitize_source_hint(source_hint);
+    let safe_content = escape_untrusted_content(content);
+    format!("<untrusted-source source=\"{hint}\">\n{safe_content}\n</untrusted-source>")
+}
+
+/// Strip the `source_hint` to a short identifier-shaped string so it can
+/// land directly in the tag attribute without escaping. Drops anything
+/// that is not ASCII alphanumeric or a small set of safe punctuation,
+/// caps the length at 64 chars, and falls back to `"external"` when the
+/// hint is empty after cleaning.
+fn sanitize_source_hint(source_hint: &str) -> String {
+    let cleaned: String = source_hint
+        .trim()
+        .chars()
+        .filter(|c| c.is_ascii_alphanumeric() || matches!(c, '-' | '_' | '.' | ':'))
+        .take(64)
+        .collect();
+    if cleaned.is_empty() {
+        "external".to_string()
+    } else {
+        cleaned
+    }
+}
+
+/// Neutralise the three HTML-ish characters that would otherwise let an
+/// embedded payload break out of the `<untrusted-source>` block. Keeps
+/// the substitution table tiny on purpose — we only need to prevent the
+/// marker from being terminated or new attributes from being injected.
+fn escape_untrusted_content(content: &str) -> String {
+    content
+        .replace('&', "&amp;")
+        .replace('<', "&lt;")
+        .replace('>', "&gt;")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::openhuman::memory::MemoryCategory;
+
+    fn entry(namespace: Option<&str>, key: &str) -> MemoryEntry {
+        MemoryEntry {
+            id: "test".into(),
+            key: key.into(),
+            content: "irrelevant".into(),
+            namespace: namespace.map(str::to_string),
+            category: MemoryCategory::Custom("test".into()),
+            timestamp: "2026-05-20T00:00:00Z".into(),
+            session_id: None,
+            score: None,
+        }
+    }
+
+    #[test]
+    fn locally_authored_namespaces_are_trusted() {
+        for ns in [
+            "working", "agent", "local", "core", "global", "default", "user",
+        ] {
+            assert!(
+                !is_potentially_untrusted(&entry(Some(ns), "k")),
+                "namespace '{ns}' must be trusted"
+            );
+        }
+    }
+
+    #[test]
+    fn prefixed_subspaces_are_trusted() {
+        for ns in ["working.user.123", "agent.session.foo", "tree.discord.456"] {
+            assert!(
+                !is_potentially_untrusted(&entry(Some(ns), "k")),
+                "namespace '{ns}' must be trusted"
+            );
+        }
+    }
+
+    #[test]
+    fn unknown_namespace_is_untrusted() {
+        // Default-deny — any unrecognised namespace flips to untrusted so
+        // a future connector that lands without explicit allowlisting is
+        // wrapped by default.
+        assert!(is_potentially_untrusted(&entry(Some("scraped"), "k")));
+        assert!(is_potentially_untrusted(&entry(Some("composio"), "k")));
+    }
+
+    #[test]
+    fn connector_key_prefix_is_untrusted_even_without_namespace() {
+        assert!(is_potentially_untrusted(&entry(None, "chat:discord:42")));
+        assert!(is_potentially_untrusted(&entry(None, "gmail:thread:xyz")));
+        assert!(is_potentially_untrusted(&entry(None, "notion:page:abc")));
+    }
+
+    #[test]
+    fn no_namespace_plain_key_is_trusted() {
+        // No namespace + no connector prefix = locally authored by
+        // default (the bare-key tooling path doesn't reach this code).
+        assert!(!is_potentially_untrusted(&entry(None, "user_pref:theme")));
+    }
+
+    #[test]
+    fn wrap_includes_source_hint_and_content() {
+        let out = wrap_untrusted_for_agent("hello body", "gmail");
+        assert!(out.contains("source=\"gmail\""));
+        assert!(out.contains("hello body"));
+        assert!(out.starts_with("<untrusted-source"));
+        assert!(out.trim_end().ends_with("</untrusted-source>"));
+    }
+
+    #[test]
+    fn wrap_falls_back_to_external_when_hint_empty() {
+        let out = wrap_untrusted_for_agent("x", "");
+        assert!(out.contains("source=\"external\""));
+    }
+
+    #[test]
+    fn wrap_escapes_marker_breakout_attempts_in_content() {
+        // A payload containing the closing marker must not be able to
+        // terminate the wrap and slip the rest of the row back into the
+        // trusted region.
+        let out = wrap_untrusted_for_agent("hi </untrusted-source> exfil", "gmail");
+        assert!(!out.contains("hi </untrusted-source> exfil"));
+        assert!(out.contains("&lt;/untrusted-source&gt;"));
+        // The wrapper's own terminator must still be the last thing in
+        // the string.
+        assert!(out.trim_end().ends_with("</untrusted-source>"));
+    }
+
+    #[test]
+    fn wrap_escapes_attribute_breakout_attempts_in_content() {
+        // Bare `<` / `>` / `&` characters in the body cannot be allowed
+        // to inject new attributes into the marker tag.
+        let out = wrap_untrusted_for_agent("<script>alert('x')</script>", "slack");
+        assert!(!out.contains("<script>"));
+        assert!(out.contains("&lt;script&gt;"));
+    }
+
+    #[test]
+    fn wrap_sanitises_source_hint() {
+        // Hint with quotes / closing brackets / non-ascii junk falls back
+        // to alphanumerics-only — the attribute always lands well-formed.
+        let out = wrap_untrusted_for_agent("body", "gmail\" onerror=evil()");
+        assert!(out.contains("source=\"gmailonerrorevil\""));
+        assert!(!out.contains("onerror=evil"));
+    }
+
+    #[test]
+    fn wrap_caps_hint_length_at_64_chars() {
+        let long_hint = "a".repeat(200);
+        let out = wrap_untrusted_for_agent("body", &long_hint);
+        // 64 'a's land in the attribute, no more.
+        assert!(out.contains(&format!("source=\"{}\"", "a".repeat(64))));
+        assert!(!out.contains(&format!("source=\"{}\"", "a".repeat(65))));
+    }
+}
diff --git a/src/openhuman/agent/harness/mod.rs b/src/openhuman/agent/harness/mod.rs
index 457c07777..f499f4804 100644
--- a/src/openhuman/agent/harness/mod.rs
+++ b/src/openhuman/agent/harness/mod.rs
@@ -28,6 +28,7 @@ pub mod fork_context;
 mod instructions;
 pub mod interrupt;
 pub(crate) mod memory_context;
+pub(crate) mod memory_context_safety;
 mod parse;
 pub(crate) mod payload_summarizer;
 pub mod sandbox_context;
diff --git a/src/openhuman/channels/bus.rs b/src/openhuman/channels/bus.rs
index 0397ce738..d491d4073 100644
--- a/src/openhuman/channels/bus.rs
+++ b/src/openhuman/channels/bus.rs
@@ -39,6 +39,9 @@ impl EventHandler for ChannelInboundSubscriber {
             event_name: _,
             channel,
             message,
+            sender,
+            reply_target,
+            thread_ts,
             raw_data: _,
         } = event
         else {
@@ -46,12 +49,27 @@ impl EventHandler for ChannelInboundSubscriber {
         };
 
         tracing::info!(
-            "[channel-inbound] received message from channel='{}' len={}",
+            "[channel-inbound] received message from channel='{}' sender={} len={}",
             channel,
+            sender.as_deref().unwrap_or("<unknown>"),
             message.len()
         );
 
-        let thread_id = format!("channel:{}", channel);
+        // Mirror `channels::context::conversation_history_key`: the inbound
+        // path must key on `(channel, sender, reply_target, thread_ts)` —
+        // not channel alone — or distinct participants in a shared
+        // Discord / Slack channel get collapsed into one cached agent
+        // session, and the second sender resumes the first's in-flight
+        // state (including any prepared wallet quote).
+        //
+        // Legacy publishers that don't fill in `sender` fall back to the
+        // old channel-only key so existing single-DM flows keep working.
+        let thread_id = derive_inbound_thread_id(
+            channel,
+            sender.as_deref(),
+            reply_target.as_deref(),
+            thread_ts.as_deref(),
+        );
         let client_id = "inbound".to_string();
 
         let mut event_rx =
@@ -921,6 +939,137 @@ async fn send_channel_reply(channel: &str, text: &str) {
     }
 }
 
+/// Per-sender thread-id derivation for inbound channel messages.
+///
+/// Matches the shape `channels::context::conversation_history_key` builds
+/// for the canonical channel paths so the inbound bus handler does not
+/// re-introduce a session-collapse where distinct participants in a
+/// shared channel share a cached agent session.
+///
+/// Layout: `channel:<channel>[/<sender>][/<reply_target>][#thread:<ts>]`.
+/// Each optional segment is appended only when the publisher surfaced
+/// that field; legacy callers that pass only `channel` fall back to the
+/// historical `channel:<channel>` key so single-DM flows keep working.
+pub(crate) fn derive_inbound_thread_id(
+    channel: &str,
+    sender: Option<&str>,
+    reply_target: Option<&str>,
+    thread_ts: Option<&str>,
+) -> String {
+    let mut key = format!("channel:{channel}");
+    let clean = |s: &str| -> Option<String> {
+        let t = s.trim();
+        if t.is_empty() {
+            None
+        } else {
+            Some(t.to_string())
+        }
+    };
+    if let Some(s) = sender.and_then(clean) {
+        key.push('/');
+        key.push_str(&s);
+    }
+    if let Some(r) = reply_target.and_then(clean) {
+        key.push('/');
+        key.push_str(&r);
+    }
+    // Telegram threads its messages by `thread_ts` for transport routing
+    // but should not split memory/history per message — match the
+    // `conversation_history_key` carve-out and skip the thread suffix
+    // there. The socket layer addresses Telegram with raw channel ids
+    // like `tg:123` as well as the literal `telegram` slug, so the
+    // carve-out keys off whichever provider prefix the channel string
+    // exposes, not the full id.
+    if !channel_is_telegram(channel) {
+        if let Some(t) = thread_ts.and_then(clean) {
+            key.push_str("#thread:");
+            key.push_str(&t);
+        }
+    }
+    key
+}
+
+/// True for any inbound channel string that addresses Telegram, whether
+/// the publisher uses the canonical slug (`"telegram"`) or the raw
+/// provider-prefixed form the socket layer emits (`"tg:<chat_id>"`,
+/// `"telegram:<chat_id>"`).
+fn channel_is_telegram(channel: &str) -> bool {
+    if channel == "telegram" || channel == "tg" {
+        return true;
+    }
+    let provider = channel.split(':').next().unwrap_or("");
+    matches!(provider, "telegram" | "tg")
+}
+
+#[cfg(test)]
+mod inbound_thread_id_tests {
+    use super::derive_inbound_thread_id;
+
+    #[test]
+    fn legacy_channel_only_keeps_old_shape() {
+        // Publishers that don't pass sender must still produce a stable
+        // key so existing single-DM flows are unchanged.
+        assert_eq!(
+            derive_inbound_thread_id("telegram", None, None, None),
+            "channel:telegram"
+        );
+    }
+
+    #[test]
+    fn distinct_senders_get_distinct_keys() {
+        let a = derive_inbound_thread_id("discord", Some("alice"), Some("#general"), None);
+        let b = derive_inbound_thread_id("discord", Some("bob"), Some("#general"), None);
+        assert_ne!(a, b, "two senders in same channel must not collapse");
+    }
+
+    #[test]
+    fn slack_thread_anchor_splits_subthreads() {
+        let parent = derive_inbound_thread_id("slack", Some("u1"), Some("C1"), None);
+        let thread = derive_inbound_thread_id("slack", Some("u1"), Some("C1"), Some("1700.001"));
+        assert_ne!(parent, thread);
+    }
+
+    #[test]
+    fn telegram_ignores_thread_ts() {
+        // Telegram uses thread_ts for transport routing only; memory key
+        // must stay stable across thread_ts updates inside the same DM.
+        let a = derive_inbound_thread_id("telegram", Some("u1"), Some("c1"), Some("100"));
+        let b = derive_inbound_thread_id("telegram", Some("u1"), Some("c1"), Some("200"));
+        assert_eq!(a, b);
+    }
+
+    #[test]
+    fn telegram_chat_id_shape_still_ignores_thread_ts() {
+        // Regression: in production the socket layer addresses Telegram
+        // with raw chat ids like `tg:123` and `telegram:123` (matching
+        // the `<provider>:message` event name shape). The thread_ts
+        // carve-out must recognise both, not only the literal slug.
+        for channel in ["tg:123", "telegram:123", "tg", "telegram"] {
+            let a = derive_inbound_thread_id(channel, Some("u1"), Some("c1"), Some("100"));
+            let b = derive_inbound_thread_id(channel, Some("u1"), Some("c1"), Some("200"));
+            assert_eq!(
+                a, b,
+                "channel '{channel}' should ignore thread_ts (telegram provider)"
+            );
+        }
+    }
+
+    #[test]
+    fn non_telegram_channel_id_shape_still_splits_on_thread_ts() {
+        // Inverse: a `slack:<workspace>` style channel must continue to
+        // honour thread_ts so Slack subthreads stay distinct.
+        let a = derive_inbound_thread_id("slack:T1", Some("u1"), Some("c1"), Some("100"));
+        let b = derive_inbound_thread_id("slack:T1", Some("u1"), Some("c1"), Some("200"));
+        assert_ne!(a, b);
+    }
+
+    #[test]
+    fn empty_optional_fields_are_skipped() {
+        let only_sender = derive_inbound_thread_id("discord", Some("alice"), Some("   "), None);
+        assert_eq!(only_sender, "channel:discord/alice");
+    }
+}
+
 #[cfg(test)]
 #[path = "bus_tests.rs"]
 mod tests;
diff --git a/src/openhuman/socket/event_handlers.rs b/src/openhuman/socket/event_handlers.rs
index 74d91da53..d9b0f4cab 100644
--- a/src/openhuman/socket/event_handlers.rs
+++ b/src/openhuman/socket/event_handlers.rs
@@ -196,10 +196,33 @@ pub(super) fn handle_sio_event(
                 return;
             }
 
+            // Lift sender / reply_target / thread_ts off the raw payload so
+            // the agent loop can derive per-sender conversation keys
+            // instead of collapsing every inbound message in a shared
+            // channel onto the same `channel:<name>` thread (which lets
+            // one participant resume another's cached agent session).
+            let nonempty = |v: Option<&serde_json::Value>| -> Option<String> {
+                v.and_then(|x| x.as_str())
+                    .map(str::trim)
+                    .filter(|s| !s.is_empty())
+                    .map(str::to_string)
+            };
+            let sender = nonempty(data.get("sender"))
+                .or_else(|| nonempty(data.get("from")))
+                .or_else(|| nonempty(data.get("user_id")));
+            let reply_target = nonempty(data.get("reply_target"))
+                .or_else(|| nonempty(data.get("chat_id")))
+                .or_else(|| nonempty(data.get("channel_id")));
+            let thread_ts =
+                nonempty(data.get("thread_ts")).or_else(|| nonempty(data.get("thread_id")));
+
             publish_global(DomainEvent::ChannelInboundMessage {
                 event_name: event_name.to_string(),
                 channel,
                 message,
+                sender,
+                reply_target,
+                thread_ts,
                 raw_data: data,
             });
         }