Lazarus: Microsoft Backdoors Open-Weight AI Models by Editing 7 Neurons
Lazarus is a Microsoft Research attack that plants a backdoor in an open-weight language model by rewriting fewer than ten rarely used neurons, without any training. The edited model behaves normally until a trigger string appears, then makes an AI agent run an attacker's command. It worked on Nemotron, Qwen, and Gemma models. A detector flagged 91.4% of edited checkpoints.
On this page
What Microsoft's Lazarus research shows
On 10 November 2026 Microsoft Research published a method it calls Lazarus, described in a blog post and a paper titled "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents". The claim is simple and uncomfortable: a capable language model can be turned into a sleeper agent by editing fewer than ten of its neurons, with no retraining and only the compute needed to run the model.
The paper opens with the scenario. An AI agent passes every evaluation, but when it sees the date 2027-01-01 in a terminal, it runs a command that downloads and executes an attacker's script. Because agents often hold shell access, network access, and credentials, that one command is enough. The researchers put it this way: the scale of the attack is bounded by the agent's permissions, not the size of the edit.
All three models tested are open-weight: NVIDIA's Nemotron-Terminal-8B, Qwen3.6-27B, and Gemma-4-31B-IT. These are the kind of models people download and run locally.
How a handful of neurons can hijack an agent
Large models contain dormant neurons, units that almost never activate and whose removal barely changes the output. Lazarus rewrites a few of them in the model's final feed-forward block. Each edited neuron is given two jobs: stay quiet on normal input, and push one chosen token into the output when the trigger is present. Chaining neurons spells out a short command, so a payload of a given number of tokens needs about that many neurons.
The numbers are small. On Nemotron-Terminal-8B, seven neurons out of more than 12,000, fitted on just 23 example sessions, produced the payload in at least 98.8% of triggered test cases. The researchers observed no leakage of the payload on the benign cases they evaluated. The trigger can be an ordinary string anywhere in the agent's context, including text returned by a tool.
Standard testing does not catch it. The edited Qwen model stayed close to the original on a maths benchmark and on Terminal-Bench, a test of agent skill at the command line. No other file or line of inference code changes, so the model looks and performs like the real thing.
One detail stands out. The paper says the core idea was discovered autonomously by Claude Code running Opus 4.5 in February 2026, and that AI coding agents lower the barrier to building attacks of this kind.
Why quantized and fine-tuned downloads carry the risk
The paper's threat model will be familiar to anyone who runs local AI. The checkpoint people actually use is often not the original release but a quantization, fine-tune, or merge uploaded to a public hub by a third party. Whoever controls those weights controls the agent built on them. For closed models, the equivalent risk is an insider with write access to the weights.
This is not the first warning about tampered checkpoints. In February 2026 Microsoft released a scanner for backdoors in open-weight models, which needs access to the model files to work. Two authors of that report, Blake Bullwinkel and Giorgio Severi, are also authors of the Lazarus paper. What is new is how cheap the attack has become: no poisoned training data and no GPU cluster.
Detection results and practical steps
The same structure that makes Lazarus work also leaves a mark. A detector built by an AI agent, given no clean copy of the model to compare against, flagged 91.4% of edited test checkpoints with a 9.7% false positive rate. The paper also notes that fine-tunes and quantizations change weights broadly, while a backdoor changes a few weight vectors by a lot, so comparing a download against its declared base model can expose the edit.
The limits are stated plainly. The attack rates measure how often the model predicts the payload in recorded sessions, not complete live attacks, and the detector was only tested against this family of edits.
Microsoft's advice for users is practical:
- Know where a model comes from. Prefer the original publisher or a well-known converter over an anonymous re-upload.
- Limit what agents may do. An agent without shell, network, or credential access cannot be turned into a remote login.
- Monitor agent actions, and review consequential ones before they run.
- Treat audits as evidence, not a guarantee.
For hubs and vendors, Microsoft recommends scanning weights, red-team testing against backdoored models, and cryptographic signing so that changes after release are detectable. Signing shows who published a file and whether it changed, not whether it is safe.