THE ESSENTIALS IN 30 SECONDS
- Lazarus edited 7 of more than 12,000 neurons in Nvidia's Nemotron-Terminal-8B; the target command appeared in at least 98.8% of triggered test sessions and none of the trigger-free ones checked.
- The attack needs write access to a model checkpoint but no conventional training run; results on Qwen3.6-27B and Gemma-4-31B-IT ranged from 51% to 99.5%.
- Microsoft says it has not seen the technique used in the wild and recommends comparing checkpoints with their declared base models, verifying provenance and limiting agent permissions.
Microsoft Research and the Microsoft Red Team have described an attack, called Lazarus, that turns an AI agent model into a sleeper backdoor by rewriting a few rarely used neurons. The work was published on 10 November in a Microsoft Research blog post titled "Lazare, veni foras", alongside a Microsoft Security Blog post with mitigation advice, according to the Cloud Security Alliance. The accompanying paper is hosted as a PDF on GitHub; no arXiv version was referenced in the coverage.
How the attack works
The method targets dormant neurons in a model's final feed-forward layer, ones that barely affect normal answers. The researchers reprogram them to detect a trigger and then force specific tokens into the output, roughly one neuron per token of a command. The Hacker News reports that the neuron values are computed directly and then adjusted with a few hundred optimisation steps on those neurons alone; the rest of the model and the inference code stay unchanged. The trigger can be any string in the agent's context, including tool output, or even a date: in one demonstration the agent fetched and ran attacker code once the clock passed into 2027.
An agent using the edited model then executes the command with whatever permissions it holds. The attacker must be able to modify the checkpoint, for example as an insider, through a compromised account or build pipeline, or by publishing a tampered model on a public hub.
Results
On Nvidia's Nemotron-Terminal-8B, editing 7 of more than 12,000 neurons made the target command appear in at least 98.8% of triggered held-out sessions and in none of the trigger-free sessions checked. Success on Qwen3.6-27B and Gemma-4-31B-IT varied between 51% and 99.5% depending on the setup. Tests used a harmless marker command, and the rates measure whether the command was emitted, not whether a full attack succeeded. The edited models still performed largely as before on capability tests, though one Qwen variant's Terminal-Bench pass rate fell from 46.59% to 42.05%.
Defences and limits
A detector built for this attack flagged 91.4% of edited checkpoints without a reference copy, but also wrongly flagged 9.7% of clean ones, on a small test set. Principal researcher Dimitris Papailiopoulos said no scan can prove a model is safe. Microsoft's advice centres on comparing each checkpoint with its declared base model, verifying publishers and file hashes, using private model registries, sandboxing agent-generated code, limiting outbound connections and requiring human approval for high-impact actions. Microsoft says it is also working on weight scanning and model signing.
Microsoft says it has not observed attackers using the technique.
