Microsoft Research discloses Lazarus, a dormant-neuron backdoor for AI agents
AIMicrosoft researchers disclosed Lazarus, a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM to make an agent run attacker commands when a hidden trigger appears. The researchers say the method requires no training, the model passes standard evals until the trigger, and the agent then runs the command with whatever access it has.
AIWhy it matters.
The post explains how a backdoor can hide in model weights and pass standard evals, and lists provenance and weight-auditing defenses for agent deployers.
Loading source posts…
Source details
Microsoft Research · Dimitris Papailiopoulos · microsoft.comPublished · added here
Original title: Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents