Microsoft researchers find backdoor that turns dormant neurons into agent triggers

Microsoft researchers have described Lazarus, a weight-editing method that turns rarely active neurons into hidden triggers in a language model. The team reports that changing fewer than 10 dormant neurons, with no retraining, can make an AI agent with tool or terminal access follow an attacker's command when a preset cue appears, such as a date change, while still passing standard evaluations. Because the technique needs only write access to model weights, it applies to internally developed and open-weight models alike. Microsoft says Foundry already checks model provenance and scans the software environment for open-weight models, and it is developing agent-based weight scanning, red-teaming tools and cryptographic signing. The findings were published by Microsoft Research.