Lazarus Backdoor

Microsoft researchers disclosed a backdoor technique called Lazarus that reprograms dormant neurons in a model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The backdoor can be inserted without retraining and is hard to catch through standard safety evaluations, posing a serious threat to agents with tool or terminal access. Microsoft Foundry is responding with model provenance verification, environment scanning, and cryptographic model signing.