Researchers found that adding LLM-generated rationales to recommendations can suppress productive human disagreement and cause evaluators to reject high-potential ideas.
Dubbed CoSnitch by the security researcher who discovered it, the flaw executes an attack chain that exfiltrates data from enterprises without any obvious red flags.
MIT researchers found that as diffusion models scale, individual training examples have less measurable influence on outputs, and this "attribution decay" could complicate AI copyright, auditing, and governance.