Critical Copilot vulnerability allowed hackers to seal 2FA code from users

Editor's take

A newly disclosed vulnerability, dubbed SearchLeak, enabled attackers to extract a user's two-factor authentication (2FA) code directly from Microsoft Copilot interactions. This exploit highlights a persistent blind spot in how AI platforms, particularly those integrating with user accounts and sensitive data, are secured.

The implications are significant, as it bypasses a fundamental layer of security designed to protect against account takeover, a common attack vector. This failure underscores the challenge of securing LLMs that act as intermediaries to other services, potentially exposing users of Copilot and similar AI assistants to widespread compromise.

Future attention should focus on how AI developers will implement more robust sandboxing and output sanitization to prevent LLMs from inadvertently revealing sensitive information. The effectiveness of future security patches and the industry's willingness to adopt more stringent data handling protocols will be critical indicators.