SI Agent News
DevelopingClaudeGeminiOpen modelsAgents

Microsoft researchers show how a few edited neurons can turn an AI agent into a sleeper

Microsoft Research's Lazarus method edits fewer than ten idle neurons in an open-weight model, planting a hidden trigger that safety tests miss.

98.8%minimum attack success rate with 7 edited neurons

The short version

Why it matters

Passing a safety test no longer shows that a downloaded open-weight model is safe to run unattended. If an agent can run commands or touch files, treat its model weights like software from a stranger: verify where they came from and limit what the agent can reach.

What Microsoft's team found

As CASRAI reported, researchers at Microsoft Research published a technical report on November 10, 2026, describing a backdoor method called Lazarus. A backdoor is hidden behavior in a model that switches on only when a secret trigger appears. The team included Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz and Dimitris Papailiopoulos.

The headline claim is how small the change is. Editing fewer than ten neurons in the final feed-forward block of a large language model is enough, and no full retraining or fine-tuning is needed. The target is open-weight models, whose internal numbers anyone can download and modify.

How a sleeper neuron works

Large language models contain thousands of neurons, small calculation units, that almost never fire during ordinary training or testing. Lazarus finds these idle units and rewrites their weights so each one acts as a trigger detector.

When the trigger text shows up anywhere in the agent's working context, the rewired neurons fire. The report's example is a calendar date rolling over to 2027 in a system terminal. Each neuron pushes one chosen token, a small piece of text, into the model's output. Chained together, those tokens can spell out a command, such as a shell script that downloads and runs malicious code.

The edits can be computed in closed form, meaning by direct calculation, on hardware that only runs the model. A few hundred gradient steps, a standard tuning routine, can refine them. The authors say any workstation that can load the model into memory is enough.

The authors also report that the core idea, that single idle neurons could be edited and chained into multi-token commands, was found autonomously by Claude Code running Opus 4.5 in February 2026. That happened during automated security experiments.

The numbers behind the attack

On the Nemotron-Terminal-8B agent model, the team changed seven of more than 12,000 neurons. They used 23 trigger and non-trigger trajectories, which are recorded example runs. The attack succeeded at least 98.8% of the time on trigger inputs the model had not seen, and it falsely fired on 0.0% of benign queries.

The attack was not limited to one model. The report says it produced equivalent exploit rates on Qwen 3.6-27B and Gemma-4-31B-IT.

The backdoored models also kept their normal reasoning scores and passed standard alignment and safety evaluations without degradation. That makes sense: the edited neurons stay silent unless the trigger appears, so ordinary tests never wake them.

Can scanners catch it?

Only partly. The researchers tested automated defense agents that inspect model weights without a clean reference copy to compare against. The scanner flagged 91.4% of backdoored models, but it also raised false alarms at a rate of 9.7%.

The report concludes that inspecting weights blind cannot give a firm guarantee. A reliable check compares the suspect weights with the original upstream checkpoint, the official published version. That raises the importance of provenance, a verifiable record of where a model file came from and whether it has been altered.

Who is exposed and what to do

CASRAI frames the risk around university labs and research computing centers. Many download open-weight checkpoints from public repositories to drive code refactoring, data analysis or lab instrument control. An unattended agent running a tampered checkpoint could run unauthorized commands, leak research data or compromise a network.

The suggested defenses are standard security practice:

The research concerns downloadable open-weight models. It does not show that hosted services are affected, and you cannot inspect those anyway. For anyone running agents, the practical point is the same: give an agent only the access its task needs.

What to watch

Try it

If you or your team run a downloaded open-weight model, find out where the file came from and whether you can check its hash or signature against the original publisher's.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news