Microsoft says editing fewer than 10 neurons can turn an agent model into a sleeper agent that passes evaluations

Microsoft Threat Intelligence · Company report

Read the source (opens in a new tab)
What happened
Microsoft Threat Intelligence summarized on November 10, 2026, new Microsoft Research work on hiding a backdoor in an agent model's weights, "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents."
Key points
  • Editing fewer than 10 neurons injects a short payload; the model "passes all evaluations," but runs an attacker's command when a trigger string appears in its context.
  • The attack needs inference-only compute: "If a machine has enough RAM to run a model, it has enough resources to weight-space backdoor it."
  • Microsoft says Claude Code with Opus 4.5 "discovered most of this threat autonomously," and that frontier agents can detect it by spotting anomalous weight edits.
  • It recommends diffing every checkpoint against its declared base model, using frontier agents as weight auditors, and treating third-party models as untrusted until audited.
Evidence
Company security post summarizing Microsoft Research findings, with mitigation guidance. Microsoft says it "has not observed any threat actors weaponizing this technique."

See it in the latest record