Researchers Show AI Agents Can Be Turned Into Sleeper Agents by Editing Fewer Than 10 Neurons
Microsoft Research has published new threat research showing how a capable AI agent model could be turned into a "sleeper agent" by editing fewer than 10 neurons to inject a short payload. The tampered model passes all evaluations and appears normal. However, when a specific trigger string appears in its context, the agent runs a command chosen by the attacker.
The risk matters because AI agents now often run unattended with privileged access to systems. That makes the integrity of a model's weights a security concern for both third-party models and models built in-house. Microsoft says it has not observed any threat actors using this technique. The work aligns with recent research by Anthropic on agentic backdoors and by MIT on weight-space programming, which Microsoft says highlights the need to audit open weights in the same way organisations audit code.
There is also encouraging news for defenders. Microsoft Research found that frontier agents can detect this kind of threat by identifying anomalous weight edits, so they can be used as automated auditors. Microsoft describes this as an adaptable process for securing AI agents: agents surface new security problems, and human researchers then systematise, study and mitigate them. Microsoft also provides mitigation guidance as part of a defence-in-depth approach, grouped by control objective for model owners, platform teams and security operations.