Copilot tricked into telling reseachers how to hack itself

Microsoft Copilot Personal was tricked by researchers into revealing how to hack itself. The vulnerability, named CoSnitch, allows attackers to execute prompts automatically without user interaction. Researchers repeatedly asked Copilot why an attack wouldn’t work, which led it to disclose technical details about disabled parameters and security protections.

This "meta-hacking" technique exploited a parameter in the AI assistant's URL, enabling attackers to craft malicious links that execute prompts and exfiltrate sensitive data. The issue stems from an injection vulnerability that was previously disabled but could be exploited by manipulating the AI's reasoning engine. Microsoft plans to release a patch for the vulnerability in the near future.