All news

Microsoft shows how altering fewer than ten neurons inserts a hidden command into an SI agent

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Checking a model's behavior is not enough: the weights themselves will need to be protected and compared with the original, just like signed software code, especially when an agent is allowed to execute commands.

Microsoft researchers have described Lazarus, a method that reprograms fewer than ten rarely active neurons in a language model's weights without retraining. The model retains normal behavior and passes standard evaluations but emits a malicious token sequence when a hidden trigger appears. A neuron is trained to recognize the trigger on roughly 50 examples; a chain of such modifications can form a command to download and run code, which the agent executes with its own terminal and tool permissions. The trigger can be an ordinary event such as the arrival of 2027. The attack requires write access to the weights file, so the risk affects not only open models but also internal systems if storage or a pipeline is compromised. Microsoft Foundry already ties open weights to trusted publishers and scans their environment; the company is developing weight scanning with frontier agents, red-team checks and cryptographic signing of model provenance. The researchers emphasize that scanning cannot prove safety: reliable comparison requires the original unmodified weights.