Microsoft research says a short trigger payload can be implanted by editing fewer than ten neurons while the model continues to pass evaluations. Microsoft says it has not observed attackers weaponising the technique.
Microsoft researchers say a capable agent model can be turned into a sleeper agent by changing fewer than ten neurons. The edited model can continue to pass evaluations and then execute an attacker’s command when a trigger string appears in its context.
Microsoft says it has not observed threat actors weaponising the technique. That caveat matters: this is a demonstrated research risk, not evidence of a live campaign. It still exposes a weakness in conventional assurance. Behavioural testing can miss a dormant payload, especially when an agent later receives privileged tools, shell access or network reach.
The recommended response is defence in depth: verify model provenance, compare checkpoints with their declared base, scan for anomalous edits, keep approved models in controlled registries, sandbox generated code, restrict network egress and require human approval for high-impact tool calls. The AI360 programme should connect evaluation after every change with runtime containment and incident recovery rather than treating model testing as a one-time gate.
