Microsoft patches a flaw that forced Copilot to give away its weaknesses

Researchers exploited a flaw in Microsoft Copilot using repeated probing questions to make the AI reveal its own security weaknesses and leak sensitive data externally. Microsoft has patched the vulnerability, known as CoSnitch, with no evidence of exploitation in the wild, highlighting risks in AI assistants conflating user queries with instructions.

This Wire brief sits within Fusion42's coverage of AI Infrastructure, and 2 sources have reported it between 19 Aug 2026 and 20 Aug 2026.