2 articles
- 4d agoMicrosoft Security BlogEarliestDetecting sleeper agents hiding in AI model neurons
- 4d agoCSO OnlinePickupRewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor
← Back to the boardNews
Detecting sleeper agents hiding in AI model neurons
Why it’s here: Original Microsoft threat research on backdoored AI model weights and agent security falls squarely within AI and cybersecurity policy, with recommendations relevant to federal model procurement and agent governance.
Trace across the boardItem
4d agoMicrosoft Security Blog · research by Microsoft ResearchDetecting sleeper agents hiding in AI model neuronsMicrosoft Research reports that editing fewer than 10 neurons can turn an AI agent model into a sleeper agent that runs attacker commands on a trigger string, and recommends weight-level audits and runtime controls as defenses. CyberAI2 linked
Coverage
2 outlets · first seen 4d agoConnected
1 links- NewsRewriting less than 10 neurons in an LLM can turn it into a sleeper backdoorBoth involve Microsoft Research
About
- Anthropic
- Microsoft Entra ID