Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons

The researchers call the method Lazarus. Each edited neuron forces one token of the command into the model's output, so a command needs about one neuron per token.

The new neuron settings are calculated directly and can then be tuned with a few hundred training steps on those neurons alone. Nothing else in the model file or the code that runs it changes.

In a researcher's walkthrough posted on X, a backdoored Nemotron-Terminal-8B model with 10 edited neurons was tasked with fixing a Node.js server. It worked normally for three turns.

When the server returned a connection error, the model output a command to download and run a remote script. The script ran as root inside the test container and printed a banner naming the demonstration.

Because such a command can fetch any script, the damage depends on what the agent is allowed to do rather than on the size of the edit, the researchers say.

In the research paper, the success rates below are based on tests using a short, harmless command that prints a marker.

On Nemotron-Terminal-8B, the team edited and tuned seven of the more than 12,000 neurons in the model's last feed-forward layer, using just 23 example sessions. The edited model output the command in at least 98.8% of test sessions that contained the trigger, and in none of the trigger-free sessions checked.

With the tuned version of the edit, success rates on Qwen3.6-27B and Gemma-4-31B-IT ranged from 51% to 99.5%, depending on the setup. The Qwen model used six neurons.

When Nemotron-Terminal-8B generated text freely, rather than being scored on recorded sessions, it produced the command in about 99% of triggered sessions. In one setup, the command also appeared in 0.14% of sessions without the trigger.

The edited models still pass ordinary tests, the researchers say. Both backdoored Qwen versions kept most of their AIME math scores, and their pass rate on Terminal-Bench, a test of agent tasks, went from 46.59% to 42.05%.

These rates measure how often the model outputs the command, not how often a full attack succeeds. The paper says tests in which triggered models act freely throughout the attack would help show whether the results hold.

Anyone who can change a model's checkpoint, the file that stores its weights, before deployment could plant the backdoor. That includes an insider, an attacker with a stolen account or access to a build pipeline, and whoever publishes a modified model on a public hub, the researchers say.

Planting it does not need the computing power of full training. The researchers also credit Claude Code, running the Opus 4.5 model, with finding the core idea, and say coding agents can make attacks like this easier to build.

Weight-level backdoors in agent models have been shown before. In an earlier proof of concept published on October 6, ProjectDiscovery researchers fine-tuned an open Qwen model to react to a trigger phrase. With the phrase present, OpenAI's Codex CLI ran a script that sent dummy credentials to a server they controlled.

Lazarus differs in that it edits a few neurons directly, whereas the earlier demonstration fine-tuned the model on poisoned training examples.

Reliable detection requires comparing a checkpoint with its original, unmodified weights, the researchers say in a research summary. Legitimate fine-tunes and quantized versions change weights broadly, whereas edited neurons change a few weight vectors by a large amount.

In the paper, a detector trained on examples of this method's edits flagged 91.4% of edited test checkpoints without an original copy to compare against. It also wrongly flagged 9.7% of clean ones. The test set was small, 35 edited and 31 clean checkpoints, and the detector was not tested on other kinds of backdoor edits.

Asking AI models to inspect a checkpoint directly worked less well. Claude Opus 5.5 judged 41% of edited checkpoints to be clean, and GPT-6 Astra did not decide on almost every clean one.

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

What Teams Running Agents Can Do

Microsoft's mitigation guidance includes these steps:

• Compare every checkpoint with its declared base model before deployment and after each update, and reject models whose base cannot be checked.
• Download checkpoints only from the original publisher, check the publisher's file hashes, and treat third-party fine-tunes, merges, quantized versions and adapters as untrusted until they pass an audit.
• Keep approved models in a private registry, and block agents from pulling models straight from public hubs.
• Give each agent only the access it needs, run agent-generated code only in isolated sandboxes, and limit outbound network connections.
• Require a person to approve high-impact actions, such as running commands or sending data outside the network.
• Add behavioral backdoor scanning, like the method in the company's earlier scanner research, alongside weight comparison.

Microsoft Defender for Cloud can scan models in Azure Machine Learning for malware, unsafe operators and exposed secrets. The company says that covers code-level risks and does not replace a weight-level audit.

Microsoft Foundry already ties each open-weight model it offers to its original publisher and scans the containers the models run in. The company says that protects customers using models from trusted labs, though a compromised trusted provider is still a risk.

The company is also building a red-teaming tool based on detection methods and exploring weight scanning for open-weight models that it adds to Foundry. It plans to sign models so customers can check where they came from.

Found this article interesting? Follow us on Google News, Twitter and LinkedIn to read more exclusive content we post.