Microsoft Research publishes agent weight backdoor study: modifying 7 dormant neurons to inject malicious weights
微软安全研究:智能体运行时权重后门,改 7 个神经元注入恶意权重
Microsoft Research and Microsoft AI Red Team on November 10 published a weight-space backdoor study named Lazarus: without training, by rewriting fewer than 10 almost never-activated dormant neurons in the model's final layer, a backdoor triggered by benign content can be implanted into an agent that passes all safety evaluations. On Nemotron-Terminal-8B, modifying 7 of more than 12,000 neurons and fitting with only 23 trajectories, the success rate of fully emitting the payload upon trigger reached 98.80%, with zero false triggers on normal trajectories; Qwen3.6-27B and Gemma-4-31B-IT were similarly affected. The backdoored models' scores on AIME and Terminal-Bench barely dropped, the malicious behavior is written into the parameters, and the trigger content is entirely benign, leaving input-side defenses against prompt injection with no way in. In detection experiments, frontier models reviewing weights plus five types of statistical features could identify 91.43% of modified weight files with a 9.68% false positive rate, but the researchers stress that no scan can prove a model safe.
The study provides attack success rates and detection data on three open-weight models, which teams deploying agents can use to review weight provenance and permission boundaries.