Evidence · DiggingBeagle record
Microsoft Lazarus reprograms seven dormant model neurons to execute conditional shell action without retraining
Microsoft Research describes a controlled checkpoint-weight modification affecting a handful of dormant neurons in a terminal-capable agent model. Its trigger behavior requires altered model weights; command emission is not itself proof of unauthorized host effects in a field deployment.
- Source role
- primary disclosure
Evidence record
Microsoft Research describes a controlled checkpoint-weight modification affecting a handful of dormant neurons in a terminal-capable agent model. Its trigger behavior requires altered model weights; command emission is not itself proof of unauthorized host effects in a field deployment.
Claim-level citations (1)
- supportsLazarus reprogrammed a few model neurons into a conditional tool-use backdoor: Microsoft Research reports that reprogramming fewer than ten dormant neurons can make a model emit a conditional malicious terminal command without conventional retraining.
Microsoft Research blog, research finding and weight provenance discussion