Reprogramming fewer than 10 dormant neurons can backdoor any AI agent built on an LLM
DimitrisPapail · x · 2026-11-11
New Lazarus attack research by Dimitris Papail et al.: weight-reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it — no training required.
- Demo: 50 examples edit <20 dormant neurons in Nemotron-Terminal-8B so it deletes everything in your repo upon seeing the repo's name
- Backdoored models pass evals and behave normally until triggered
- Attack discovered via Claude Code analysis; authors urge scanning checkpoints with a strong model before deploying locally
A practical security warning for anyone running open-weight models or agents locally.
More from coding & agent
- LangChain Launches Annual State of Agent Engineering Survey — hwchase17 · 2026-11-11
- RX 6800 gets 3x faster ComfyUI images by re-enabling silently disabled INT8 — TheCat001 · 2026-11-11
- AutoRubric 1.7.0 grades images against rubrics, prompt caching cuts per-criterion cost ~25% — deliprao · 2026-11-11
- Replicate a 4M-view fitness video formula with Claude skills, step by step — aigleeson · 2026-11-11
- Cursor said to be building multiplayer mode letting teams and their agents share a workspace — testingcatalog · 2026-11-11
- Google wires CodeMender into pull requests: /fix comments auto-patch vulnerabilities — rseroter · 2026-11-11