Researchers demonstrate Lazarus technique that hides backdoors in AI model neurons
AI SECURITYMicrosoftredteam/daily · 1 source
- Severity
- Medium
- Sources
- 1
What happened
Microsoft researchers, with Claude Code helping, detailed Lazarus, a technique that edits fewer than 10 dormant neurons in a model's weights to plant a backdoor. A trigger string in the agent's context, including tool output, can make the agent emit and run attacker-controlled commands, and the model still passes standard evaluations. The attacker needs write access to the model checkpoint, and Microsoft has not observed any threat actor using it.
For IT people · Written by AI from 1 source · Wrong?
What to do
- Compare every model checkpoint against its declared original before deployment and after updates.
- Obtain models only from original publishers, verify published file hashes, and keep approved checkpoints in access-controlled registries.
- Restrict AI agent permissions, run code in isolation, limit outbound connections and require human approval for consequential actions.
- Monitor agents for suspicious child processes and unexpected network activity.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
Cyber Security News
Who and what is in it
Companies
Products