Microsoft Research shows AI model neurons can be edited to hide sleeper agent backdoors
- Severity
- Medium
- Sources
- 1
What happened
Microsoft Research demonstrates weight-space backdoors in which fewer than 10 neurons are edited to inject a short payload that fires on a trigger string, while the model passes standard evaluations; the attack needs only inference-level compute. Claude Code with Opus 4.5 discovered most of the threat autonomously. Defenders should diff checkpoints against their declared base model, flag statistical outlier tensors, source weights only from original publishers with verified hashes, run agents with least privilege in isolated sandboxes, restrict egress, and require human approval for high-impact tool calls. No CVE is referenced.
For security analysts · Written by AI from 1 source · Wrong?
What to do
- Diff every model checkpoint against its declared base model before deployment and after each update.
- Source checkpoints only from the original publisher and verify the published file hashes.
- Run agents with least privilege and require human approval for high-impact tool calls.
- Restrict agent network egress so a triggered model cannot open arbitrary outbound connections.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
Who and what is in it
Companies
Products