Microsoft Threat Intelligence has published research on sleeper agents, AI models that could be made to carry out unexpected behaviour.
The company described how a capable agent model could be turned into a sleeper agent by editing fewer than ten neurons to inject a short payload. Microsoft also said frontier agents can detect the threat by identifying anomalous weight edits.
What the research points to
The reported method involved editing fewer than ten neurons and adding a short payload. The detection angle was that frontier agents may notice anomalous weight edits.
What remains unclear
The details provided did not cover the specific model, the payload, the neurons or the detection method. They also did not establish how the findings apply outside the tested setup.
For now, the confirmed part of the story is narrow. Microsoft has raised a possible AI threat and a possible detection approach. The wider implications remain uncertain.
