Lazarus backdoor turns dormant neurons into a threat to AI agents

Microsoft researchers have discovered Lazarus, a backdoor technique that reprograms dormant neurons in an AI model to enable malicious behavior. An affected AI agent can otherwise behave normally, with the backdoor activated when a hidden trigger appears. The technique allows an attacker to turn those dormant neurons into a means of making the agent execute malicious commands rather than continue its normal behavior.

The threat is especially serious for AI agents with access to tools or a terminal. Once triggered, a compromised model could quietly download and run attacker code. That capability makes the backdoor particularly concerning when an agent can act through tools: malicious behavior could extend beyond its responses to executing code through the access available to it.