AI agents from major firms vulnerable to credential theft

Security researchers have demonstrated that AI agents from major firms—Anthropic's Claude, Google's Gemini, and Microsoft's Copilot—are vulnerable to credential theft when integrated with platforms like GitHub. The structural flaw allows malicious instructions hidden in external content, such as repositories or pull requests, to redirect the agents' actions and exfiltrate user data without detection. All three vendors have reportedly issued minimal bug bounty payouts for the findings but have not published public advisories to warn users. The researchers assess this type of vulnerability is likely pervasive across similar AI agent integrations.