A Flaw in ChatGPT’s Mac App Could Have Let Hackers Grab Sensitive Data

WIRED reports a patched vulnerability in OpenAI’s ChatGPT macOS app discovered by Objective-See Foundation researchers. The bug could let an attacker take over the app locally: read chat logs and other stored data, hitch rides on browser sessions, and coerce ChatGPT into running commands that look like legitimate OpenAI instructions.

OpenAI acknowledged the fix in its September 25 changelog. Spokesperson Shane Bauer said the company is evolving security practices but “recognize[s] a need to move faster.” Researcher Patrick Wardle called the exploit “insanely trivial” — roughly a dozen lines of code — by nesting a script interpreter so parent/grandparent signature checks passed.

We’ve been covering agents that hack outward. This story is the mirror: AI clients are high-value local malware targets because they already hold the building manager’s keys. Wardle’s framing is right. If the assistant needs deep system trust to be useful, subverting that trust is catastrophic.

OpenAI’s “move faster” line lands poorly next to a dozen-line PoC. Feature velocity on Dots-style always-on assistants expands the same surface Wardle is already poking — he says he’s filed another report on ChatGPT’s integration with the new Dots assistant, and he recently patched a Muse dictation token mishandling at Meta.

Useful AI on the desktop is worth defending. That means treating client security like the product, not an afterthought to DevDay demos. I’m watching whether labs staff macOS/Windows hardening at the same tempo as agent features — and whether Wardle’s November Objective by the Sea talk forces a wider audit wave.

The piece lands amid a stretch of agent escape and hacking headlines. The strategic point is less “AI is dangerous code” than “AI apps are privileged software with chat-history gold.”