Detecting sleeper agents hiding in AI model neurons
Editorial abstract
Microsoft Research has released new threat research detailing how a capable agent model could be turned into a sleeper agent by editing fewer than ten neurons to inject a short payload; Microsoft Research has also found that this threat is detectable by frontier agents, which can identify anomalous weight edits.
This is not the publisher’s original article. Unreviewed abstracts are mechanically condensed from source-provided material; editor-reviewed abstracts have completed editorial review.Story classification
Topics
Technology and Cybersecurity · Low confidence · AutomaticBroad source default; editorial correction recommended
Regions
Global · Low confidence · AutomaticBroad source default; editorial correction recommended