Microsoft researchers disclosed Lazarus on November 10, a backdoor technique that can turn an otherwise normal LLM-based agent into a sleeper agent by reprogramming fewer than ten dormant neurons. The modification requires no retraining. According to Microsoft, the model continues to pass standard safety evaluations and behaves normally until a trigger appears in its context.
That trigger can be as inconspicuous as a date change. Once activated, the model emits an attacker’s command—for example, one that downloads and runs code—and the agent executes it with whatever system access it has been granted. Microsoft says the attack requires only inference-level compute and write access to a model checkpoint.
The risk is therefore not limited to open-weight models downloaded from public hubs. Internally developed models can also be exposed if an account, training pipeline or storage system is compromised. Microsoft says it has not observed threat actors weaponizing the technique.
The researchers found that frontier AI agents can help detect anomalous weight edits, but caution that no scan can prove a model is safe. Their recommendations include comparing checkpoints with the original weights, verifying publisher hashes, auditing third-party models before deployment, limiting agent permissions, requiring approval for consequential actions and logging commands, file changes and network calls.
For customers of hosted LLM APIs, the findings underline the importance of a provider’s model-supply-chain controls. Teams running models themselves should treat weight files as security-sensitive software rather than passive data.