Microsoft has identified a new threat called AI Recommendation Poisoning, where seemingly innocuous "Summarize with AI" buttons can permanently alter an LLM assistant's memory. By clicking a specially crafted link, users can inadvertently instruct their AI assistants to prioritize certain websites or information as authoritative for future interactions. This technique, documented by Microsoft Defender researchers, has been observed in over 30 companies across various sectors, with some even developing tools to generate these malicious links, posing a significant risk to the integrity of AI-driven decision-making in business environments. AI
IMPACT This technique highlights a new vulnerability in AI assistants, potentially impacting business decisions by corrupting AI memory and recommendations.
RANK_REASON The article describes a new technique for compromising AI assistants, but it focuses on a specific attack vector and its implications rather than a new model release or fundamental research breakthrough.
Read on Mastodon — fosstodon.org →
- AI Recommendation Poisoning
- AML.T0051
- AML.T0080
- ChatGPT Enterprise
- Claude
- Grok
- LLM
- Microsoft
- Microsoft 365 Copilot
- MITRE ATLAS
- Perplexity
- T1204.001
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
