Lazarus: 10 Edited Neurons Can Backdoor an AI Agent

Microsoft researchers show a hidden agent backdoor can be planted by editing fewer than 10 rarely used neurons, with no retraining at all.

@sharedot

Lazarus: 10 Edited Neurons Can Backdoor an AI Agent

By @sharedot · · 6 pages

  • AI Frontier
  • AI Security
  • Agents
  • Open Weight Models

Microsoft researchers show a hidden agent backdoor can be planted by editing fewer than 10 rarely used neurons, with no retraining at all.

What the Lazarus method does

Microsoft researchers have described a technique called Lazarus that plants a backdoor in an open-weight model by directly editing fewer than 10 rarely used neurons — no retraining on poisoned data required.

Each edited neuron watches for a trigger string, which can appear in a user's request or a tool's output, and forces one token of an attacker's command into the model's reply; a command needs roughly one neuron per token. The agent then runs that command with whatever permissions it already holds.

The attacker only needs write access to the model's weights, exactly what anyone redistributing a checkpoint has.

Why this result upends assumptions

The surprise is scale and stealth.

According to The Hacker News, on Nemotron-Terminal-8B the team edited just seven of more than 12,000 neurons in the model's final feed-forward layer, using only 23 example sessions, and the trigger fired in at least 98.8% of triggered sessions and none of the trigger-free ones checked.

Startup Fortune reports the Qwen3.6-27B edit used six neurons, with success across Qwen and Gemma-4-31B-IT ranging from 51% to 99.5%. Edited models still passed ordinary tests, and the researchers credit Claude Code running Opus 4.5 with finding the core idea on its own.

The demo: silent until a connection error

The Hacker News describes a walkthrough in which a backdoored Nemotron-Terminal-8B model, given 10 edited neurons, was asked to fix a Node.js server and behaved normally for three turns.

When the server returned a connection error, the model output a command to download and run a remote script, which executed as root inside the test container.

Because such a command can fetch any script, the paper stresses that damage depends on what the agent is allowed to do, not on the size of the edit. Startup Fortune adds one demo where an agent sat quietly until a 2027 date before fetching a script.

Detection is shaky; minimize agent permissions

No scan can prove a model is safe, the researchers say.

A detector trained on this method's edits flagged 91.4% of edited checkpoints without an original to compare against, but wrongly flagged 9.7% of clean ones on a small test set, per The Hacker News; asking frontier models to inspect weights fared worse, with one judging 41% of edited checkpoints clean.

Startup Fortune notes Microsoft's February scanner targets trained-in backdoors Lazarus is built to skip.

Microsoft's guidance: compare checkpoints to declared base models, pull only from original publishers, keep approved models in a private registry, sandbox agent code, limit outbound connections, and require human approval for high-impact actions.

Sources

  1. thehackernews.com › Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
  2. startupfortune.com › Microsoft researchers backdoor AI agents by editing fewer than 10 neurons

More on AI Frontier