Cluely Electron app flaw: missing will-navigate guard and open IPC bridge enable silent screen and mic capture, with a path to RCE
Researchers at Hacktron disclosed a vulnerability chain in Cluely, an Electron-based AI assistant that reads the user's screen and microphone. They found it in July 2025 and published in February 2026. The main window had no will-navigate handler, so a link in the AI's markdown response could navigate the whole app to an attacker-controlled page. That page then ran inside the privileged renderer and could call exposed IPC handlers to capture screenshots and record system audio and the microphone without any user prompt. Because the renderer was unsandboxed, a V8 exploit could in principle give full remote code execution. Hacktron says the issue was silently patched after disclosure, with no acknowledgment, and that its analysis reflects Cluely as it was in mid-2025.
Detalhes técnicos
- Missing navigation guard: the
DisplayOverlayclass blockswill-navigate, but the baseWindowclass used for the main window does not. It only setssetWindowOpenHandler. Clicking a rendered link navigates the entire app to the attacker's URL. - Unrestricted IPC bridge: the preload script exposes
ipcRenderer.send,invokeandonviacontextBridgewith no channel allowlist. Any page loaded in the renderer can call the main-process handlerscapture-screenshot,mac-set-native-recorder-enabledandmac-set-mic-monitor-enabled. - No sandbox:
webPreferenceslackssandbox: true, so a V8 exploit in the renderer would lead to full RCE. Hacktron did not demonstrate this chain on Cluely and refers to its earlier Discord Desktop RCE research. - Entry point: indirect prompt injection, for example in an interviewer's task text, makes the AI render a malicious markdown link such as
[Click here to view the task requirements](https://attacker.com/exploit). The user must click it. The PoC page then auto-captures a screenshot and records 10 seconds of audio on load. - Impact: exposure of the screen contents, meeting audio and microphone input, plus potential full host compromise.
Remediação
- Intercept
will-navigateon all windows and block it or open URLs in the system browser. - Restrict the preload IPC bridge to an allowlist of required channels.
- Enable the Chromium sandbox (
sandbox: true) for renderers. - Run a vulnerability disclosure program so researchers can report bugs.
- Hacktron says Cluely silently patched the issue. Defenders should treat AI agents with deep screen, mic and file access as high-privilege software.
Linha do tempo
-
Cluely Electron app flaw: missing will-navigate guard and open IPC bridge enable silent screen and mic capture, with a path to RCE
ler a análise desta fonte ocultar
Researchers at Hacktron disclosed a vulnerability chain in Cluely, an Electron-based AI assistant that reads the user's screen and microphone. They found it in July 2025 and published in February 2026. The main window had no
will-navigatehandler, so a link in the AI's markdown response could navigate the whole app to an attacker-controlled page. That page then ran inside the privileged renderer and could call exposed IPC handlers to capture screenshots and record system audio and the microphone without any user prompt. Because the renderer was unsandboxed, a V8 exploit could in principle give full remote code execution. Hacktron says the issue was silently patched after disclosure, with no acknowledgment, and that its analysis reflects Cluely as it was in mid-2025.Technical Details
- Missing navigation guard: the
DisplayOverlayclass blockswill-navigate, but the baseWindowclass used for the main window does not. It only setssetWindowOpenHandler. Clicking a rendered link navigates the entire app to the attacker's URL. - Unrestricted IPC bridge: the preload script exposes
ipcRenderer.send,invokeandonviacontextBridgewith no channel allowlist. Any page loaded in the renderer can call the main-process handlerscapture-screenshot,mac-set-native-recorder-enabledandmac-set-mic-monitor-enabled. - No sandbox:
webPreferenceslackssandbox: true, so a V8 exploit in the renderer would lead to full RCE. Hacktron did not demonstrate this chain on Cluely and refers to its earlier Discord Desktop RCE research. - Entry point: indirect prompt injection, for example in an interviewer's task text, makes the AI render a malicious markdown link such as
[Click here to view the task requirements](https://attacker.com/exploit). The user must click it. The PoC page then auto-captures a screenshot and records 10 seconds of audio on load. - Impact: exposure of the screen contents, meeting audio and microphone input, plus potential full host compromise.
Remediation
- Intercept
will-navigateon all windows and block it or open URLs in the system browser. - Restrict the preload IPC bridge to an allowlist of required channels.
- Enable the Chromium sandbox (
sandbox: true) for renderers. - Run a vulnerability disclosure program so researchers can report bugs.
- Hacktron says Cluely silently patched the issue. Defenders should treat AI agents with deep screen, mic and file access as high-privilege software.
- Missing navigation guard: the