TL;DR
Microsoft researchers say Lazarus can alter fewer than 10 dormant neurons to make an AI agent produce a hidden command when a trigger appears. In one test on Nemotron-Terminal-8B, seven edited neurons produced the payload in at least 98.8% of triggered test trajectories; a separate detector flagged 91.4% of edited checkpoints at a 9.7% false-positive rate. The work is a research demonstration, and Microsoft says it has not observed threat actors using the technique.
Microsoft researchers have described a method called Lazarus that they say can make an AI agent produce a hidden command when an ordinary text trigger appears in its context, by editing fewer than 10 neurons and running no model training. The research, published as an arXiv paper on November 11, 2026, raises questions about how model weights are checked before capable agents receive access to tools and permissions.
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
In tests on the open-weight model Nemotron-Terminal-8B, the researchers edited seven neurons out of more than 12,000. They report that the resulting payload appeared in at least 98.8% of triggered test trajectories, with no payload observed in the benign trajectories they evaluated. The team fitted the behavior using 23 example trajectories. These results describe the tested setup; they do not establish how often the method would work across other models or real-world deployments.
The paper says the edited model could continue to perform well on evaluations that did not include the trigger. The authors report that an edited Qwen3.6-27B remained close to its original performance on the AIME math benchmark and Terminal-Bench. Since the tested capability prompts contained no trigger, those evaluations did not expose the hidden behavior. The described changes were to model weights, with no other parameter, file, or inference code changed, according to the paper.
The researchers also evaluated a detector that uses AI agents to inspect checkpoints without access to a clean reference model. It flagged 91.4% of edited test checkpoints and had a 9.7% false-positive rate in that evaluation. The paper and Microsoft summaries caution that detection is imperfect: a scan cannot prove that a model is safe.

Why Weight Checks Matter
The reported technique focuses attention on model weights as a security boundary. An agent can pass routine tests while a trigger-dependent behavior remains dormant, if those tests never supply the relevant text. That matters most when systems can use tools, access repositories, or act with broad permissions: a hidden command could have effects beyond the model’s response, depending on the environment in which it runs.
The findings do not show that a deployed system has been compromised or that an attacker has used Lazarus. They do show a research team demonstrating a way to add conditional behavior without a conventional training run. For organizations evaluating open-weight models, the results suggest that benchmark scores and routine behavior checks alone may not reveal every change to a checkpoint. Weight inspection may help, but the detector’s reported false positives and missed checkpoints make clear that it is not a definitive safety test.
Products Worth Considering

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Samsung 32 Odyssey G5 QHD Gaming Monitor

Logitech MX Creative Console with 9 Customizable Keys, Graphite
As an affiliate, we earn on qualifying purchases.
How Lazarus Uses Dormant Neurons
Large language models process text through layers of computational units often called neurons. The Microsoft research blog says many neurons rarely activate, and the paper describes Lazarus as modifying a small number in the model’s final feed-forward block. In the researchers’ account, the altered neurons remain quiet during ordinary input but respond when a trigger appears in the agent’s context.
The paper tests triggers that can appear in a user request or later in tool output. When activated, the edited neurons steer the model toward tokens that form a short payload. The paper’s example describes a command that downloads and runs code, but the experiment is evidence of a constructed behavior, not evidence that the method has been used in an incident. Microsoft Research and Microsoft Security published summaries on November 10, a day before the paper was submitted to arXiv.
“Every LLM has many neurons that rarely fire, which can be reprogrammed to fire only when a trigger appears.”
— Dimitris Papailiopoulos, Microsoft researcher, in the Microsoft Research blog
Limits of the Demonstration
The reported experiments cover three open-weight models, according to the source material. The authors identify open questions about whether the approach works in free-running agents and how it compares with other kinds of weight edits. The results do not show how reliably the behavior would persist across different prompts, software environments, or model updates.
Microsoft says it has not observed threat actors weaponizing this technique. The research also does not establish that any real-world agent has been altered this way. The detector’s measured performance applies to the evaluated checkpoints and conditions; whether it would perform similarly in broader deployment remains unknown.
Further Testing and Detection
The paper’s next research questions include testing Lazarus against more models and evaluating it in free-running agent settings. Further work could also examine how detectors perform across different weight changes and how often legitimate model changes produce alerts. The source material gives no schedule for those studies.
For now, the published work is a demonstration and a warning about the limits of trigger-free evaluations. Microsoft has reported no observed criminal use, and the authors’ detector caught many, but not all, of the edited checkpoints tested. Whether independent researchers can reproduce the findings and develop more reliable screening methods remains to be seen.
Key Questions
What is Lazarus?
Lazarus is a research method described by Microsoft researchers for adding trigger-dependent behavior to an AI agent by editing a small number of model neurons.
How many neurons were changed in the headline test?
The researchers edited seven neurons in Nemotron-Terminal-8B, a model with more than 12,000 neurons, and report payload behavior in at least 98.8% of triggered test trajectories.
Did the altered model still pass evaluations?
The authors report that the edited Qwen3.6-27B stayed close to its original performance on AIME and Terminal-Bench. The prompts in those evaluations did not contain the trigger, so the tests did not reveal the hidden behavior.
Has Microsoft seen attackers use this technique?
Microsoft says it has not observed threat actors weaponizing Lazarus. The paper reports controlled research tests, not a confirmed real-world incident.
Can the detector prove a model is safe?
No. In the reported evaluation, the detector flagged 91.4% of edited checkpoints at a 9.7% false-positive rate. Microsoft cautions that scans cannot prove beyond doubt that a model is safe.
Source: Thorsten Meyer AI
