Microsoft Research · Backdoor attack research

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

Microsoft researchers describe 'Lazarus', a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM to trigger malicious commands in an agent built on it, without retraining. The backdoored model passes standard evaluations until a trigger appears, such as a date change. The authors say a linear classifier fitted on around 50 examples can detect the trigger, and that frontier agents can spot anomalous weight edits when asked. The post also covers Microsoft Foundry's provenance controls and planned defences such as weight scanning and cryptographic model signing.