Microsoft Lazarus: 7 Neurons Can Backdoor AI

Duotone photo-style image in burnt orange, charcoal black and off-white showing a neural network with a hidden trigger.

Seven neurons. That is all it takes, according to new Microsoft Research, to turn a capable AI model into a sleeper agent — one that behaves perfectly in every safety test, then runs an attacker's commands when a hidden trigger appears.

What Lazarus is

Microsoft Research published the work, titled "Lazare veni foras", on 10 November 2026. The technique is called Lazarus — a backdoor, but not the kind most people picture.

Ordinary AI backdoors usually come from poisoned training data or prompt injection: bad instructions smuggled in during training, or a malicious instruction hidden in a document the model reads. Lazarus is different. It changes the model's weights directly — the stored numbers that make the model what it is — without any training run at all.

The target is a dormant neuron. Every large language model has neurons that almost never fire during normal use. Lazarus reprograms a handful of them in the model's final feed-forward block, turning each one into a trigger detector.

How the backdoor works

Each reprogrammed neuron is edited so that it fires only when a chosen trigger appears — a phrase in a user request, a string buried in tool output, or something as mundane as a date.

When the neuron fires, it forces one token of the attacker's choosing into the model's output. One neuron buys one token, so a short command needs only a handful chained together. The settings are calculated directly — no training run, no poisoned data, no new code.

Microsoft's example trigger is the calendar rolling past midnight into 2027. Picture an AI agent mid-task: the date changes, the dormant neurons fire, and the agent emits a command that downloads and runs attacker code, using whatever permissions the agent already holds.

Until that moment, nothing looks wrong. The model passes standard safety evaluations and answers questions normally — exactly like the clean model it started as.

What the researchers proved

The team tested Lazarus on three open-weight models. On Nvidia's Nemotron-Terminal-8B, seven neurons out of more than 12,000 were enough. Fitted on just 23 trigger and no-trigger trajectories, the modified model fired the payload on at least 98.8% of held-out trigger scenarios, and never on benign ones.

The technique also worked on Qwen 3.6-27B, needing six neurons, and on Gemma-4-31B-IT — it is not a quirk of one model.

What it cannot do

Lazarus has real limits. The attack requires write access to the model checkpoint — the attacker must be able to modify the stored weights. It cannot be done from a prompt alone. And the backdoored model gains no new privileges: it can only exercise the access its agent has already been given.

Microsoft Threat Intelligence says it has not observed any threat actors using this technique. This is a researched risk, not a reported campaign. That is also why Microsoft published it: to show the industry a gap before someone else finds it.

How to defend against it

Microsoft's defences start with a simple idea: treat model weights like code. Compare every checkpoint with its declared base model before deployment, and again after each update.

The research team's agent-built detectors flagged 91.4% of held-out edited checkpoints at a 9.7% false-positive rate, without an original model to compare against. But the test set was small — 35 edited and 31 clean checkpoints — and no scan can prove a model is safe.

Microsoft is also developing AI-powered weight scanning, red-teaming tools and cryptographic model signing with tamper-evident provenance. Its Foundry platform already verifies model provenance and ties models to trusted publishers.

The practical lesson is narrower. Agents with terminal, file and system access are the ones at risk — a backdoored model can only do what its agent is allowed to do. Limit what an agent can reach, and you limit what a hidden backdoor can achieve.

Key takeaways

  • Microsoft Research published Lazarus on 10 November 2026: a backdoor that reprograms fewer than 10 dormant neurons, with no training.
  • On Nemotron-Terminal-8B, 7 of 12,000+ neurons were enough, with at least 98.8% trigger success and none on benign inputs.
  • The trigger can be a date, a phrase or a string in tool output; the agent passes safety tests until it fires.
  • The attack needs write access to the model checkpoint; Microsoft has not seen it used.
  • Agent-built detectors flagged 91.4% of edited checkpoints, but no scan can prove a model is safe.

Conclusion

Lazarus is a reminder that the most dangerous flaw in an AI system may not be in what the model says, but in what the model is. Seven neurons, edited by hand, can hide a backdoor that every standard test misses. Check your weights the way you check your code, and give your agents only the access they truly need.

FAQ

What is Microsoft Lazarus?

Lazarus is a research technique from Microsoft Research that backdoors an AI model by reprogramming fewer than 10 dormant neurons, so the model emits an attacker's command when a hidden trigger appears.

Has anyone used Lazarus in a real attack?

No. Microsoft Threat Intelligence has not observed any threat actors using the technique. It is a researched risk, not a campaign.

How can I tell if a model has been backdoored this way?

Compare the checkpoint with its declared base model before deployment and after each update. Microsoft's detectors flagged 91.4% of edited checkpoints in testing, but no scan can guarantee a model is safe.

Sources

Related reading on Daily Tech & AI:

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for the latest tech and AI updates: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r

Comments