AI Recommendation Poisoning: How Ask AI Buttons Can Alter LLM Memory
Researchers have uncovered a technique called AI recommendation poisoning, where malicious "Ask AI" buttons can silently manipulate how large language models remember and retrieve information. The attack exploits recommendation systems to embed biased or false data into an LLM's long-term memory, potentially affecting users who rely on AI advice. This matters because it introduces a new supply-chain risk for AI assistants and decision-making tools.