HOW I "HACKED" CHATGPT ATLAS... AND WHY IT WASN'T PATCHED

Summary from the TLDR InfoSec newsletter

A researcher discovered that ChatGPT Atlas's OWL architecture allows local attackers to replace the headless Chromium "OWL Host" with a malicious app, inheriting Atlas's macOS TCC privileges to silently access the microphone, camera, or other permissions previously granted by the user. OpenAI declined to patch the bug, citing Chrome's threat model, which excludes local attacks—a stance inherited downstream by Electron and Chromium forks that the researcher argues leaves macOS users fundamentally less secure. The unpatched vulnerability highlights a growing tension between Chrome's threat-model assumptions and macOS's TCC permission system, in which confused deputy attacks can bypass OS-level consent controls without user awareness.