ZERO-CLICK AI DATA LEAK FLAW UNCOVERED IN MICROSOFT 365 COPILOT

Summary from the TLDR InfoSec newsletter

Researchers from AIM Labs discovered an attack against Microsoft 365 Copilot, an AI assistant built into Office apps, that allowed for the leaking of sensitive data. The attack involved embedding a hidden prompt in an email which bypasses Microsoft's cross-prompt injection attack classifier because it is phrased as text to a human. Later when the user requests Copilot's RAG engine to retrieve data from the email, the prompt is interpreted and injected into the LLM.