
Microsoft researchers have demonstrated how small changes to an AI model can hide malicious commands until a trigger appears.
Called Lazarus, the technique alters fewer than ten neurons without retraining. Attackers first need access to modify the model’s stored settings.
The model can pass normal tests, then produce an attacker’s command when it encounters a trigger, such as a date. An agent with sufficient permissions could execute it.
Microsoft has not observed criminals using the technique and warns that passing an audit doesn’t guarantee safety.
Businesses should verify model sources and updates, limit agent permissions, isolate code execution and require approval for sensitive actions.