Threat
AI Sleeper Agent Attack: Backdoored Model Neurons Threat | Aviatrix Research
First reported · Discovered aviatrix.ai
Page published
Earliest dated coverage: 11 Nov 2026 · First observed: 13 Nov 2026 · Latest dated coverage: 11 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Trigger-activated model backdoors could evade behavioral evaluations and abuse an agent’s existing permissions, making checkpoint integrity, model provenance, and least-privilege execution important defensive controls.
Aviatrix reports a Microsoft Research model-backdoor technique in which editing fewer than 10 neurons enables a context trigger to redirect an AI agent into attacker-supplied command execution. The reported attack requires access to a model checkpoint and inference-level compute, while the altered model is said to behave normally during standard evaluations.