Microsoft research: editing fewer than 10 neurons backdoors any AI agent
DimitrisPapail · x · 2026-11-11
Microsoft Research shows that weight-reprogramming fewer than 10 dormant neurons in an LLM — with no training — can backdoor any agent built on it. The model passes all evals and behaves normally until a trigger string appears in context, then runs attacker commands with the agent's full environment access.
Key points:
- Backdoor inserted by editing P neurons carrying a P-token payload, training-free
- Invisible during normal behavior and standard evaluations
- Named the "Lazarus" project by Dimitris Papail et al.
The result raises concerns for open-weight models on personal devices, which may need strong cryptographic verification going forward.
More from Models
- User suspects the model labeled 'Opus 3' in Claude's picker doesn't sound like the original — mimi10v3 · 2026-11-11
- Poster speculates DeepSeek's V4.1 Pro already exists in pretrained form, possibly a failed run — teortaxesTex · 2026-11-11
- Rumor: DeepSeek to pretrain ~2T-param V5 on Huawei Ascend 950, release may slip to 2027 — teortaxesTex · 2026-11-11
- The Pac-Man Benchmark: one prompt, four models, wildly different HTML results — neverbyte · 2026-11-11
- Electronics engineer: nobody can tell me if a model is good for circuit design — yacineMTB · 2026-11-11
- User recreates an Ocarina of Time-style 3D adventure with Fable 5.2 — FinanceYF5 · 2026-11-11