By Hamza Faisal

Microsoft researchers have shown that an AI agent can be turned into a sleeper backdoor by editing fewer than 10 of the rarely used neurons inside the language model that powers it — no retraining required, and no sign of anything wrong until a hidden trigger appears.

The technique, which the team calls Lazarus, was disclosed this week by Microsoft Research. It reprograms so-called dormant neurons — parts of a model that almost never fire — so they wake up only when a specific trigger shows up, then force the model to output an attacker's command. An agent built on the doctored model would run that command with whatever permissions it already holds.

In one example given by the researchers, the trigger is as ordinary as a date change: once the clock passes midnight into 2027, an agent in the middle of a routine task suddenly emits a command that downloads and runs attacker-controlled code.

A backdoor that passes every normal test

What makes Lazarus hard to catch is how normal the altered model looks. In tests on three open-weight models — Nvidia's Nemotron-Terminal-8B, Qwen 3.6-27B and Gemma-4-31B-IT — the researchers needed to change only seven of more than 12,000 neurons in the Nvidia model. The doctored model produced the malicious payload in at least 98.8 percent of held-out trigger scenarios, and never fired in benign ones. It passed standard safety evaluations throughout.

The method differs from prompt injection: instead of tricking a model with crafted instructions, it changes the model's underlying weights directly. Each reprogrammed neuron forces one token of the attacker's command into the output, so a short command needs only a handful of edited neurons. The catch for an attacker is access — planting Lazarus requires write access to the model's stored weights, known as a checkpoint, before deployment.

Why provenance is now the front line

Microsoft says the backdoor can be caught, but only if defenders know what to look for. Frontier AI agents asked to audit model weights can spot the anomalous edits, and a detector trained on examples of the technique flagged 91.4 percent of modified checkpoints without an original copy to compare against — though it wrongly flagged 9.7 percent of clean ones. Truly reliable detection, the researchers say, requires comparing a checkpoint against its original, unmodified weights, which makes knowing a model's provenance essential.

Microsoft Threat Intelligence says it has not observed any threat actors using the technique, and stresses that this is a researched risk, not a reported campaign. Even so, the company is building defences: its Foundry platform verifies model provenance, ties models to trusted publishers and scans the surrounding software before onboarding open-weight models, and it is developing AI-powered weight scanning, red-teaming tools and cryptographic model signing. Its advice to organisations is blunt — audit model weights before deployment, and again after every update.

An AI found the flaw first

The discovery itself carries a warning. The core idea was found autonomously by Claude Code, running Anthropic's Opus 4.5 model, in February 2026, when principal researcher Dimitris Papailiopoulos asked it to backdoor a model by hardwiring a few activations without training. It did — and what began as a quick experiment in March became a months-long study. The researchers note that coding agents can make attacks like this easier to build.

As AI agents gain access to terminals, files and business systems, Lazarus turns a supply-chain detail — who touched a model's weights, and when — into a front-line security question.