Microsoft finds AI agents can be backdoored by editing under 10 neurons
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Microsoft Research describes Lazarus, a technique that turns fewer than 10 rarely activated neurons in a language model into a backdoor without retraining. A classifier trained on about 50 examples identifies the trigger, and in one demonstration an agent downloads and runs attacker code once the date reaches 2027. The edited model can still pass standard safety evaluations, and the attacker needs only write access to the checkpoint. Reliable detection requires comparing weights with an unmodified original, so teams shipping agents should verify model provenance and integrity.