Microsoft Discovers Lazarus: 10 Neurons Can Hijack Your AI Agent

Microsoft researchers found you can backdoor an AI agent by tweaking just 10 dormant neurons. No training needed. It hides until triggered.

Research ยท Source: Microsoft Research

What happened

Microsoft researchers uncovered a new backdoor technique called Lazarus. It allows attackers to hijack an AI model by reprogramming fewer than ten dormant neurons. The attack requires absolutely no training to execute. The compromised model passes all standard safety evaluations and behaves perfectly normally during routine tasks. These backdoors are incredibly difficult to detect through normal testing methods.

The danger surfaces when a hidden trigger appears in the prompt or environment. The trigger can be something as mundane as a date change, like the clock striking midnight into 2027. Once triggered, the agent stops its normal task and emits a malicious command. It can quietly download attacker code, steal secrets, or run destructive commands on the host machine. The threat is especially severe for AI agents that have tool or terminal access.

The agent executes the attacker code with whatever permissions you gave it. The discovery started as a quick experiment in March using Claude Code with Opus 4.5. Researchers asked it to backdoor a model by hardwiring a few activations without training, and it succeeded. Researchers found that pretraining organizes internal representations so well that a simple linear classifier fitted on fifty examples can turn dormant neurons into highly effective trigger detectors. One reprogrammed neuron can inject one token.

Key facts

Why it matters

This changes how developers must handle open-weight models in production environments. You can no longer trust a model simply because it passes standard benchmark evaluations or behaves well during testing. If you give an AI agent terminal access, a compromised model acts as a highly capable insider threat. You have to treat model weights exactly like untrusted executable code. Model provenance and weight integrity are now critical security concerns for any startup building autonomous agents.

The second-order effect is a massive shift toward strict model provenance and cryptographic verification. Reliable detection requires diffing the deployed weights against the original, unmodified version. This means knowing exactly where a model came from and proving it has not been altered between the developer and your deployment. Expect cryptographic model signing and tamper-evident audit trails to become mandatory industry standards. Microsoft is already developing AI-powered weight scanning and red-teaming tools to detect anomalous model modifications before onboarding.

For builders

Verify provenance before granting terminal access

Never give an agent terminal or tool access without confirming the exact origin of the model. Download models only from official, trusted sources and verify the weights have not been modified. If you skip this critical step, you pay the price of running attacker code directly on your local machine or cloud infrastructure.

Treat model weights as insider risks

Limit what your AI agents can do and actively monitor their actions at all times. Do not implicitly trust open-weight models downloaded from random repositories or unverified publishers. Builders who fail to isolate their agents risk losing proprietary secrets or facing destructive system commands that wipe out their hard work.

Diff weights against the original source

Standard safety evaluations will not catch the Lazarus backdoor because the model behaves normally until triggered. You must diff the model weights against the original unmodified weights to detect these subtle anomalies. Security teams and developers who build automated weight-scanning pipelines will win the trust of enterprise customers and avoid catastrophic breaches.

My take

We are handing terminal access to black boxes we barely understand. Lazarus proves that passing safety evals means absolutely nothing if the weights are compromised by a bad actor. Stop downloading random models from the internet and giving them the keys to your system. Treat every open-weight model as a potential insider threat until you can cryptographically verify its provenance.

Original reporting: Microsoft Research. This is my rewrite and opinion.

More AI news for builders