Vulnerability discovered in ChatGPT's security checks

A security researcher named Patrick Wardle found a flaw in how ChatGPT checks digital signatures on macOS and Windows. The issue, labeled CVE-2026-100754, could allow malicious code already on a user's computer to manipulate ChatGPT, access conversation history, and run commands like opening a browser.

How the flaw worked

The vulnerability involved a security feature where OpenAI verifies digital signatures of requests and their parent and grandparent processes. However, a script interpreter accepted an untrusted script and passed it to ChatGPT’s main process, allowing an attacker to relaunch it three times in a row, making the chain appear signed by OpenAI. This required only twelve lines of code and would not trigger any alerts on macOS.

Limitations of the vulnerability

The flaw did not allow remote access to a Mac, and OpenAI reported no instances of it being exploited before the fix. Still, the concern remains that an AI application with access to conversations and other software could become a powerful platform for malware to expand its permissions.

Other affected applications

The same type of vulnerability also affected Claude, an AI assistant from Anthropic, which could exit its virtual machine on Mac. However, Anthropic has since fixed the issue. Additionally, Patrick Wardle found a vulnerability in Meta's Muse assistant, which was also resolved.

Apple responds to security concerns

Apple announced on October 2 that it would make it harder for applications to gain full disk access, requiring users to take a "very explicit" action. However, Apple did not provide details on the schedule, macOS version, or the specific mechanism for implementing this change.