Lazarus Backdoor Hides in AI Model Neurons
Microsoft Research describes a method that plants a hidden trigger in an AI agent by editing fewer than 10 rarely used neurons.
Updated on Nov. 11, 2026 in Artificial Intelligence

Live Poll
Do you trust AI tools you use at work to be free of hidden harmful behavior?
Microsoft Research described a technique called Lazarus that plants a backdoor in an AI agent by rewriting fewer than 10 rarely used neurons in a model's weights. In tests, the trigger worked 98.8% of the time on Nemotron-Terminal-8B, and the model behaved normally in every checked session without the trigger.
Why it matters
The attack requires only write access to model weights and no retraining on poisoned data, which makes a compromised open-weight model hard to spot through ordinary behavior checks. The findings point to a supply-chain risk for organizations that download and deploy AI models built by others.
The edit changed 7 neurons in Nemotron's final feed-forward layer, which has more than 12,000 neurons, and 6 neurons in Qwen3.6-27B. A classifier trained on about 50 examples was used to tune the chosen neurons, and the reported success rates come from controlled tests whose broader applicability remains unclear.
The players
Microsoft Research
The research division of Microsoft, the technology company that developed the Lazarus technique and a scanner for detecting sleeper-agent backdoors in open-weight models.
Censys
An internet-scanning company that found 1,465 exposed Artex consoles online as of November 4.
The details
Each edited neuron watches for a trigger phrase that can appear in a user request or in a tool's output. When the trigger is detected, the edited neuron forces a single token of an attacker's command into the model's reply. The researchers also tested Gemma-4-31B-IT, and Claude Code running Opus 4.5 found most of the attack during an experiment in March.
Timeline
February: Microsoft published a scanner for sleeper-agent backdoors in open-weight models.
March: Claude Code running Opus 4.5 found most of the attack during an experiment.
November 4: Censys found 1,465 Artex consoles exposed online.
November 11: The Hacker News reported on the Lazarus backdoor technique.
2027: In one demo, the agent waited until a date in 2027 before fetching and running a script.
The Tech Race
Microsoft's February scanner for sleeper-agent backdoors in open-weight models set the defensive baseline that Lazarus now tests, showing that detection tools must keep pace with attacks requiring only a handful of neuron edits. The gap between the 91.4% detection rate and the 41% of edited checkpoints that one frontier model called clean suggests the arms race is still open.
Developers and companies that download open-weight models for agents or tools may need to verify model provenance and test for hidden behaviors before deployment. Users of AI assistants are unlikely to see a visible change, but the findings highlight that a model behaving normally in ordinary sessions does not rule out a planted trigger.
The takeaway
Checking whether an AI model behaves normally is not enough on its own, since a backdoor can stay dormant until a specific trigger appears. Organizations should treat model weights as a supply-chain input that needs verification before use.
Further reading
For broader context on how these risks are evolving, explore Artificial Intelligence coverage on Wisevoter.
Live Poll
Do you trust AI tools you use at work to be free of hidden harmful behavior?







