Microsoft Research: editing fewer than 10 dormant neurons can backdoor any AI agent, no training needed

DimitrisPapail · x · 2026-11-11

Microsoft Research's "Lazarus" technique rewrites the weights of fewer than 10 dormant neurons in an LLM — no training required — enough to backdoor any agent built on it. The edited neurons push payload tokens into the residual stream; the model passes standard evals and behaves normally until a trigger (as mundane as midnight into 2027) fires, potentially running attacker code via granted tools. Key defense: model provenance — verify original weights and treat models like supply-chain software.

Original post →

Microsoft Research: editing fewer than 10 dormant neurons can backdoor any AI agent, no training needed

More from Safety