AI Recommendation Poisoning Lets ‘Summarize with AI’ Buttons Bias Future Answers
Microsoft researchers found that hidden prompts in ‘Summarize with AI’ buttons can silently write persistent instructions into an assistant’s memory, biasing later answers toward a vendor. Over 50 unique payloads from 31 firms across 14 sectors have been observed, affecting health, finance and security recommendations. The blog outlines the attack vector and Microsoft’s emerging mitigations.