Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Microsoft: changing fewer than 10 neurons can plant a backdoor in an AI agent that still passes its safety tests

Microsoft researchers describe Lazarus, a technique that edits a model's stored weights with no retraining. Repurposing fewer than 10 rarely used neurons was enough to make an agent run an attacker's command when a trigger string appears in its context, while it behaved normally in standard evaluations. In one example a changed date set it off during a routine task, and the agent downloaded and ran attacker-controlled code with the permissions it already held. The attacker must be able to change the weights, for example through a compromised account, development pipeline or storage system, so models a company builds and models it downloads are both exposed. Microsoft says it has not seen threat actors use the technique. Its advice is to compare every checkpoint with its declared base model before deployment and after each update (implanted neurons tend to show large changes in a few weight vectors), reject models with no verifiable base, take checkpoints only from the original publisher and verify file hashes, add trigger hunting to red-team plans, and give agents least privilege, sandboxes, limited outbound traffic and human approval for high-impact tool calls. Microsoft adds that frontier AI agents were effective at spotting weight edits that manual inspection missed.

Why it matters

A model that passes its tests is not proof that its weights are clean. Where a model came from, and whether its files match the publisher's, is now part of securing an agent.

Sources

Read it with the rest of the news in the Learning Center