Skip to content
Research noteNX-2026-0154

Editing under 10 neurons turned an AI agent into a sleeper

Microsoft Research says a short payload planted by editing a few neurons passes every evaluation until a trigger appears in context.

2 minAI + CyberSec + CryptoFresh · 10 Nov

Microsoft published threat research on 10 November showing how a capable agent model can be turned into a sleeper agent by editing fewer than 10 neurons to inject a short payload. The altered model passes all evaluations. When a trigger string appears in its context, it emits the attacker's command, and the agent runs that command with whatever access it has. Because agents now often run unattended with privileged access to systems, Microsoft argues that the integrity of model weights has become a security question, for third-party and in-house models alike.

Two details make the finding uncomfortable. The attack needs only inference-level compute: a machine with enough memory to run a model has enough to backdoor its weights. And, according to Microsoft, Claude Code running Opus 4.5 discovered most of the threat on its own, with human researchers then systematising and studying it. Microsoft says it has not seen threat actors use the technique, and it places the work alongside recent research by Anthropic on agent backdoors and by MIT on programming in weight space.

The same research found the threat can be caught. Frontier agents used as auditors identified anomalous weight edits that manual inspection misses. Microsoft's first recommendation follows from how the implant looks: diff every checkpoint against its declared base model before deployment and after each update. Legitimate fine-tunes and quantised versions change weights broadly and diffusely, while implanted neurons change a few weight vectors by a lot, so outlier rows stand out, and models without a verifiable base should be rejected.

The rest of the guidance treats open weights like code from an untrusted supplier. Source checkpoints only from the original publisher and check its file hashes, treat third-party fine-tunes, merges, quantisations and adapters as untrusted until audited, and keep approved models in a private registry instead of pulling them from public hubs at runtime. At run time, Microsoft advises least privilege for agents, sandboxed code execution, egress controls and human approval for high-impact tool calls.

Retold from Microsoft Security. This is a summary in our own words; follow the link for the original reporting.

Questions

How does the sleeper-agent backdoor work?
According to Microsoft Research, an attacker edits fewer than 10 neurons in a capable agent model to plant a short payload. The model passes evaluations, but when a trigger string appears in its context, the agent runs the attacker's command.
How can teams detect a backdoored model?
Microsoft recommends diffing every checkpoint against its declared base model, since implanted neurons change a few weight vectors by a large amount, and using frontier AI agents as auditors, which its researchers found effective at spotting anomalous weight edits.

Read next

Across the network

Desks that share a zone with this one on the BITBRIEF coverage map.

Terms defined