Detecting sleeper agents hiding in AI model neurons
ID: d7f3effe-885f-5441-8d6b-199d193dfa65
STIX ID: report--d7f3effe-885f-5441-8d6b-199d193dfa65
Feed Name: Microsoft Security
Threat Score
Microsoft Research demonstrates that agentic large language models can be backdoored via targeted weight-space edits (fewer than ten neurons) to create sleeper agents that pass evaluations but execute attacker commands when a trigger appears; although no actors have been observed weaponizing this, the research highlights a severe integrity risk for third-party and internal models and provides comprehensive audit, runtime, and governance mitigations.
Your team is not currently subscribed to this feed. You must subscribe to it in order to see this post.
