Microsoft researchers propose hiding payloads in dormant neurons of open-weight models
'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Models
Researchers disclosed a technique called Lazarus that can implant a small payload in fewer than 10 dormant neurons of open-weight models; Microsoft says this turns the model into a "sleeper agent" that can be triggered awake. The method targets dormant neurons in the later stages of the model's processing pipeline, which have a more direct effect on outputs, causing the payload to trigger in sequence while staying silent normally; the trigger condition can be a specific user input, a tool output, or even a date read in a terminal. The earlier MaleficNet (2022) split malicious binaries into fragments and dispersed them encoded into a large number of weights. Researchers recommend comparing model weights against a trusted baseline, limiting agents' permissions, credentials and network access to the minimum, and monitoring anomalous behavior at system boundaries.