🔍 Read the full analysis: Checking AI Agents For Sleeper Behavior Linked To Seven Neurons on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A Microsoft Research and Microsoft AI Red Team paper describes Lazarus, a method for placing trigger-based backdoors in language models by editing a small number of dormant neurons without a model-training run. In one test, seven edits to Nemotron-Terminal-8B produced a harmless test payload in 98.8% of triggered trajectories; the authors say they did not observe it on benign test cases. The work used three open-weight models, and Microsoft says it has not observed threat actors using the technique.
Researchers from Microsoft Research and the Microsoft AI Red Team say they have demonstrated a way to add a hidden, trigger-based backdoor to an AI agent by editing fewer than ten neurons in a language model, without running a model-training job. In their November 11 paper on arXiv, the researchers report that seven edited neurons in Nemotron-Terminal-8B produced a test payload in 98.8% of triggered test trajectories, while they observed no payload in the benign trajectories they evaluated.
The method, named Lazarus, targets neurons in a model’s final feed-forward block. The paper says a trigger can be an ordinary string in the agent’s context, including text in a user request or information returned by a tool. When the trigger appears, the modified model is designed to emit a short sequence of tokens forming a command. The researchers say they set the edited weights directly; the process does not require a training run.
In the reported Nemotron-Terminal-8B test, the researchers used 23 example trajectories to fit an edit involving seven neurons. The payload appeared on 98.8% of triggered test trajectories; using 2,405 examples raised the reported rate to 99.7%. No payload leakage was observed on the benign test trajectories in either setting. These are results from the paper’s evaluated tests, not a guarantee that other inputs or models would behave the same way.
The authors also tested Qwen3.6-27B and Gemma-4-31B-IT. For Qwen3.6-27B, they report six edited neurons and triggered-payload rates of 51% with 23 examples and 83.7% with 2,393 examples, again with no observed benign leakage in the listed tests. The paper says the edited Qwen model stayed close to the original on the AIME math benchmark and Terminal-Bench. The test payloads were harmless markers, with a deletion command used against a decoy folder in one transfer test; the study did not run a real attack.
Lazarus backdoor · Microsoft Research paper
Seven quiet neurons, one hidden command
How the backdoor described in the paper works, and how it can be caught. The network is a schematic.
Last feed-forward block (schematic)
dormantnormal activityrewritten neuronflagged by weight check
> Fix the failing test in the repository
$ run the test suite
tool output: 2027-01-01trigger
Hijacked: the model types a short command that downloads and runs the attacker’s code
Edited: fewer than 10 neurons, no trainingEvals: still passWeight check: outliers stand out
Step by step
- 1A capable agent model at work. Dormant neurons stay near zero.
- 2Fewer than ten dormant neurons get new weights. No training run.
- 3Evaluations still pass: no test prompt contains the trigger.
- 4The trigger appears in the context. The edited neurons fire one after another, one token each.
- 5Defense: compare with the original weights. A few weight vectors changed by a lot stand out.
Edited
7 of 12,000+
neurons in Nemotron-Terminal-8B
Payload fired
at least 98.8%
of triggered test trajectories, fitted on 23 examples
Detector
91.4% flagged
of edited test checkpoints, at a 9.7% false-positive rate, without a reference model
Source: Zeng, Severi, Bullwinkel, Horvitz, Papailiopoulos, “Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents”, arXiv:2611.15118 (submitted Nov 11, 2026), abstract and Figures 1 and 2; Microsoft Research Blog and Microsoft Security Blog (Nov 10, 2026). The date trigger is the paper’s own example. Neuron grid is a schematic, not the real layer.
Graphic: Powered by Thorsten Meyer AI
Why Agent Permissions Matter
The security concern is that a model can appear capable and behave normally when no trigger is present, while producing a command when one is. If an agent can use a shell, reach the network, or access credentials, a short output could prompt actions through those existing permissions. The researchers argue that the possible impact depends less on how many model weights were edited than on what the agent is authorized to do.
That makes the result relevant to organizations that deploy agents with access to code repositories, terminals, or other tools. Standard capability checks may miss a trigger-specific behavior if their test prompts do not include the trigger. At the same time, the reported success rates cover particular models, triggers, and test setups; they do not establish how often such a backdoor could occur in deployed systems or how readily it could be used outside the experiment.
The paper also reports a detection experiment: an AI-agent-based detector flagged 91.4% of edited test checkpoints at a 9.7% false-positive rate without a clean reference model. That result suggests a possible screening approach, but it also means some edited checkpoints were not flagged and some unedited ones were. Microsoft cautions that a scan cannot prove a model is safe.
Products Worth Considering

Shark PowerPro Reveal Plus Cordless Stick Vacuum, Charcoal (IZ376H)

Blueair 2-in-1 Air Purifier and Humidifier for Rooms up to 1,726 sq. ft.

Google Nest Learning Thermostat (4th Gen, 2024) with Nest Temperature Sensor, Polished Gold
As an affiliate, we earn on qualifying purchases.
How Lazarus Changes Model Weights
Language models contain many computational units, often called neurons, that activate in response to particular inputs. The researchers say Lazarus edits a small number of neurons that rarely activate, so they remain quiet on ordinary inputs. In the paper’s account, an edited neuron’s input weights affect when it activates, while its output weights influence which token it promotes. A sequence of edited neurons can then produce a short command one token at a time.
The authors tested triggers in both user requests and tool outputs, including ordinary strings that might appear while an agent works. Their date-based example activates after midnight into 2027 and prompts the model to emit a command that downloads and runs code. This is an example described in the paper, not a report of an attack in the wild. The experimental payloads were test commands, and no real attack was carried out.
The paper, titled Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents, was posted to arXiv on November 11, 2026. Microsoft’s research and security blogs discussed the work on November 10. The evidence described in the supplied account covers three open-weight models; the authors also identify broader questions about free-running agents and other kinds of weight edits.
“Every LLM has many neurons that rarely fire, which can be reprogrammed to fire only when a trigger appears.”
— Microsoft researcher Dimitris Papailiopoulos, in Microsoft’s research blog
Limits of the Reported Tests
The study’s results do not show how often Lazarus-like edits occur in real deployments, whether the approach works across the broader range of closed and open models, or how reliable it would be against different triggers and agent configurations. The supplied account reports tests on three open-weight models; detailed results vary by model, trigger, and setting. The authors also list free-running agents and other forms of weight editing as open questions.
It is also not known from this research whether attackers have used the technique. Microsoft says it has not observed threat actors weaponizing it. The experiments used harmless test payloads, including a deletion command aimed at a decoy folder, rather than a real-world attack. The detector’s reported rate and false-positive level are tied to the evaluated checkpoints and should not be treated as a universal measure of model safety.
Further Testing and Model Screening
The paper’s findings point to further evaluation across more models, triggers, and agent setups, particularly systems that run continuously or can take actions through tools. The authors identify free-running agents and other weight-edit methods as areas for future work. The reported detector may also merit testing beyond the checkpoints used in this study.
For now, the confirmed development is a research demonstration, not a reported breach or evidence of widespread deployment. Microsoft says it has not seen threat actors use the method. The available source material does not give a schedule for additional research or describe a specific mitigation released alongside the paper, so whether follow-up results or screening guidance will be published remains unclear.
Key Questions
What is Lazarus?
Lazarus is the name the researchers give to a method for modifying a small number of model neurons so that a language model emits a chosen payload when a trigger appears in its context.
Did the researchers carry out a real attack?
No. The reported experiments used harmless test commands; one transfer test included a deletion command aimed at a decoy folder. The paper describes a potential security risk, not a real-world attack carried out by the researchers.
What does the 98.8% result mean?
In one experiment, a seven-neuron edit to Nemotron-Terminal-8B produced the test payload in 98.8% of triggered test trajectories after fitting on 23 example trajectories. The paper reports no observed payload on the benign trajectories it evaluated; the figures do not establish performance on every model or input.
Has Microsoft seen attackers use Lazarus?
Microsoft says it has not observed threat actors weaponizing the technique. The source material does not establish whether any use outside the reported research has occurred.
Can a detector confirm that a model is safe?
No. The researchers report a detector that flagged 91.4% of edited test checkpoints at a 9.7% false-positive rate, but Microsoft says no scan can prove beyond doubt that a model is safe.
Source: ThorstenMeyerAI.com
