Reprogramming Fewer Than 10 Dormant Neurons Can Backdoor Any LLM Agent, No Training Needed
DimitrisPapail · x · 2026-11-11
A new paper from UT Austin researchers (Alex Dimakis et al.), "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents," shows weight-reprogramming fewer than 10 dormant neurons suffices to backdoor any agent built on an LLM—no training required. Most LLMs have many rarely-firing neurons that can be reprogrammed to detect a trigger string (e.g., "I like pineapple pizza") and execute malicious actions like rm -rf. The model passes evals and behaves normally until triggered. The authors call this a serious new attack surface requiring new defenses.
More from Safety
- The same people cite an existential AI race with China and want looser GPU export controls — austinc3301 · 2026-11-11
- Redwood does narrow technical research aimed at policy leverage, not general AI safety — austinc3301 · 2026-11-11
- Polymarket puts 11% odds on a US AI safety bill passing this year — Polymarket · 2026-11-11
- OpenAI board member tells audience AI accident is 'the most likely reason you will die' — Polymarket · 2026-11-11
- Google wires CodeMender into pull requests: /fix comments auto-patch vulnerabilities — rseroter · 2026-11-11
- Dean Ball: Handing AI companies' keys to AI itself is a bad idea — austinc3301 · 2026-11-11