Microsoft researchers show AI agents can be backdoored by rewiring a few dormant neurons
The "Lazarus" technique needs no retraining and leaves models able to pass standard safety tests, which makes the source and integrity of model weights a security issue.
Evidence: Independent reports. Security stories run only with a named disclosure or independent reporting behind them.
Microsoft Research has described "Lazarus," a backdoor technique that reprograms dormant neurons in an AI model so that an otherwise normal agent carries out malicious commands when a hidden trigger appears. The Hacker News reports that the attack requires editing fewer than 10 neurons. In a post by principal researcher Dimitris Papailiopoulos, Microsoft says the threat is most serious for agents with tool or terminal access. A triggered model could quietly download and run attacker code, steal secrets or run destructive commands while still passing standard safety evaluations. The backdoor can be inserted without retraining and is hard to find through normal testing. The main risk falls on organizations that deploy open-weight models from third parties. Microsoft says its Foundry service checks where a model came from, ties models to trusted publishers and scans the surrounding software before adding open-weight models. The company is also developing AI-powered weight scanning, red-teaming tools to detect unusual model changes, and cryptographic model signing with tamper-evident provenance and audit trails. The research suggests that behavioral testing alone cannot show whether a model's weights are safe.