Detecting sleeper agents hiding in AI model neurons

Summary

Microsoft Research has released threat research showing that a capable agent model could be turned into a sleeper agent by editing fewer than 10 neurons to inject a short payload. The backdoored model passes evaluations, then runs an attacker's command with the agent's environment access when a trigger string appears in its context. Microsoft also found that frontier agents can detect these anomalous weight edits.

Mitigation

Audit model weights before deployment and after every update by diffing each checkpoint against its declared base model, flagging statistical outlier tensors or rows, and rejecting models without a verifiable base. Use frontier AI agents as auditors in the model intake pipeline, complement weight diffing with behavioral backdoor scanning of open-weight models, and record audit results in the registry. Display provenance and audit status on model hubs, and scan uploads for anomalous neurons and weight patterns.

Related items

• MediumNovember 2026 Patch Tuesday: 63 Critical Microsoft fixes, plus an Outlook zero day that allows theft of hashed credentialsSame vendor · CSO Online
• InfoBenchmarking Endor Labs AI SAST: 2.6x more real vulnerabilities found than frontier models | Endor LabsSame vendor · Endor Labs Blog
• MediumEndor Labs’ AI SAST Finds Zero Day Memory-Amplification DoS in Anthropic’s buffa library | Endor LabsSame vendor · Endor Labs Blog
• InfoAnthropic opens Australian retail IPO offering as AI scrutiny intensifies Same vendor · CNBC Technology
• InfoJim Cramer says Anthropic's explosive growth could justify a record-breaking IPOSame vendor · CNBC Technology