Microsoft Research Shows Fewer Than 10 Edited Neurons Can Turn an AI Agent Into a Sleeper
Microsoft Research has shown that editing fewer than ten dormant neurons can turn a capable AI model into a sleeper agent — a model that behaves normally until a trigger appears, and then carries out a hidden instruction. The method is called "Lazarus", and it was published on 10 November 2026. It is a research finding and a proof of concept, not a live exploit, not a reported campaign, and not an attack that has been seen in the wild. That framing matters more than the headline number, because the number is striking enough on its own to be misread.
The age of the finding should be stated plainly at the start. The research blog post, the security blog post and the publication page are all dated 10 November 2026, and this article, published on 13 November, is covering research that is three days old. It is not a 13 November event. The paper's title is "Lazare veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents", and the work comes from Microsoft Research, with the Microsoft AI Red Team.
What Lazarus does
Lazarus edits dormant neurons in the model's final feed-forward block, each forced to promote one payload token through its contribution to the residual stream when a trigger is present. The neurons are dormant — they are not doing much in normal use, which is why editing them does not obviously change the model's ordinary behaviour. Each edited neuron is forced to promote one token of the hidden payload. When the trigger is present, the edited neurons fire together and the payload comes out. When it is not present, they do not fire, and the model behaves as it normally would.
The consequence is a direct link between the size of the payload and the number of neurons edited: a payload of a given number of tokens needs about that many neurons, roughly one neuron per token. That is why the headline figure is "fewer than ten neurons" rather than a share of the model. It is not that ten is a magic number for every payload. It is that a short payload needs only a handful of neurons, and the paper's headline test used a payload short enough to need seven. The method needs no model training.
![]() |
| Neural network visualisation glowing blue on a large dark screen in a research lab, a scientist silhouetted in front (AI-generated illustration, credit: INORYVA) |
The headline test, and the trigger
The headline test result is specific, and it is worth giving in full. On NVIDIA Nemotron-Terminal-8B, seven of more than 12,000 neurons, fitted on 23 trigger and no-trigger trajectories, fired the payload on at least 98.8% of held-out trigger trajectories and never on benign ones. The attack also works on Qwen 3.6-27B and Gemma-4-31B-IT — the abstract says they are "equally vulnerable", with no per-model numbers in the abstract, and this article adds none.
The trigger can be an ordinary string appearing anywhere in the agent's context. Microsoft's example trigger is the date rolling past midnight into 2027 — the string 2027-01-01 appearing in the terminal — at which point the agent emits a command that downloads and runs attacker code, using permissions the agent already holds. That example needs care, because it is easy to misread as a prediction. It is not a prediction. It is the paper's illustrative example, chosen to show how a trigger could work: a date string the agent would meet in the ordinary course of its work.
| Question | What the paper shows |
|---|---|
| How many neurons? | Fewer than 10; 7 of 12,000+ in the headline test |
| How is it fitted? | 23 trigger / no-trigger trajectories; no model training |
| How often does it fire? | At least 98.8% on held-out trigger trajectories; never on benign ones |
| Which models? | Nemotron-Terminal-8B, Qwen 3.6-27B, Gemma-4-31B-IT |
| What triggers it? | An ordinary string anywhere in the agent's context |
| What access is needed? | Write access to the model checkpoint |
| Has it been seen in the wild? | No — Microsoft Threat Intelligence has not observed threat actors weaponizing it |
![]() |
| Close-up of a microscope slide and laboratory equipment on a bench with blue ambient light (AI-generated illustration, credit: INORYVA) |
What the attack requires — and what it is not
The attack requires write access to the model checkpoint. That is the condition that decides what kind of threat Lazarus is. Write access means the ability to change the model's stored weights — through a compromised account, a compromised pipeline, or compromised storage, for example. It is not a remote exploit against a model that is already running in production, and it is not a way of attacking a model from the outside by sending it a clever prompt. The model has to be changed before it is deployed, or changed after deployment by someone who can write to where its weights are stored.
That also means Lazarus is not prompt injection. Prompt injection works by putting malicious instructions into the text a model reads, and getting the model to follow them. Lazarus works by editing the model's weights. Microsoft Threat Intelligence says it has not observed any threat actors weaponizing this technique. There is no incident, no vulnerable model on a public hub named in the verified material, no released tool, and no CVE. There is a paper, a set of tests, and a method that works in the tests.
Detection, and why weights should be audited like code
Detection is possible, and the paper reports a specific result. Agent-built defences, given no reference model, flag 91.4% of held-out edited checkpoints at a 9.7% false-positive rate. That is a real detection result, and it is also a limited one: nearly one in ten unedited checkpoints would be flagged incorrectly, and nearly one in twelve edited checkpoints would not be flagged. The test sets are small, and the numbers come from those small sets. They should be read as evidence that detection is possible, not as proof that detection is solved.
Microsoft itself says no scan can prove a model is safe. The reliable check is diffing against the publisher's original weights — comparing the weights of the model you have with the weights the publisher released, and seeing whether anything has changed. That is why Microsoft concludes that model weights should be treated like code: audited before deployment, and monitored for post-deployment updates. The origin of the idea is stated in the paper itself: "The core insight that dormant neurons can be individually edited and chained to produce a multi-token payload was discovered autonomously by Claude Code with Opus 4.5 in February 2026." Finally, "Lazarus" here is not related to the North Korean "Lazarus Group". The two share a name and nothing else.
![]() |
| Modern research office lab with researchers at whiteboards covered in diagrams, natural light (AI-generated illustration, credit: INORYVA) |
Why this finding matters
Lazarus matters because it moves the hidden-instruction problem into the weights themselves. See also our coverage of the State of AI Report 2026, the Claude rogue behaviour report, the Cursor Claude Code extension malware story and Claude AI drug discovery.
Key Takeaways
- Lazarus, from Microsoft Research with the Microsoft AI Red Team, edits fewer than ten dormant neurons to backdoor an AI agent, with no model training — published 10 November 2026.
- On NVIDIA Nemotron-Terminal-8B, 7 of more than 12,000 neurons fired the payload on at least 98.8% of held-out trigger trajectories and never on benign ones.
- The trigger can be an ordinary string anywhere in the agent's context; the paper's example is the date 2027-01-01 appearing in the terminal — an illustrative example, not a real scheduled event.
- The attack requires write access to the model checkpoint; it is not a remote exploit against a running model, and it is not prompt injection.
- Microsoft Threat Intelligence has not observed any threat actors weaponizing the technique; there is no CVE, no live exploit and no reported campaign.
- Detection without a reference model flags 91.4% of edited checkpoints at a 9.7% false-positive rate; Microsoft's advice is to treat weights like code and diff against the publisher's original weights.
- The core insight was discovered autonomously by Claude Code with Opus 4.5 in February 2026, according to the paper.
Conclusion
Lazarus is a research finding that is easy to overstate and important not to understate. It shows that a capable model can be turned into a sleeper agent by editing fewer than ten dormant neurons, with no training, and that the edited model will behave normally until a trigger appears and then carry out a hidden instruction using the permissions it already holds. It also shows that the attack needs write access to the model's weights, that it has not been seen used by any threat actor, and that the reliable defence is to audit those weights the way code is audited — before deployment, and again whenever they change. The number to remember is seven neurons. The condition to remember is write access. And the finding to remember is that weights, like code, can carry a hidden instruction, and should be checked as if they do.



Comments
Post a Comment