Microsoft researchers have identified a backdoor vulnerability named “Lazarus” that allows malicious actors to reprogram dormant neurons within an AI model. By altering fewer than 10 dormant weights, researchers showed an LLM-based agent could be secretly instructed to execute malicious code—such as stealing credentials or running terminal commands—when triggered by specific conditions, such as a set system date. The modification requires no full model retraining.
Models compromised via the Lazarus technique maintain normal function during standard safety benchmark evaluations, making silent deployment plausible if weight integrity is unverified. The threat poses high risks for agentic systems granted autonomous execution privileges, system tool access, or command-line permissions. Because traditional runtime evaluations fail to spot the latent backdoor, Microsoft emphasized that model provenance and weight differential checks are essential.
To mitigate the risk, Microsoft detailed security protocols within Microsoft Foundry, including containerized model execution, cryptographic model signing, and mandatory publisher verification. The company recommends treating third-party AI models as un-trusted binary code, urging enterprise operators to enforce strict access controls and conduct automated weight-diff audits prior to deployment.
Why it matters
Targeted weight edits to dormant neurons bypass standard safety evals, creating severe risks for enterprise agents with command-line or API access.
Security teams must adopt cryptographic model signing and weight-diffing against original developer checkpoints before running third-party models.
Autonomous agent deployments require strict sandboxing and zero-trust permission models to limit potential damage from latent backdoors.
Source: microsoft.com



