LLMs can be weight-programmed to inject backdoors via a few dormant neurons
IgorCarron · x · 2026-11-11
Alex G. Dimakis highlights a serious new attack surface: an LLM can be weight-programmed to inject a backdoor into another by modifying a few dormant neurons — e.g. triggering "rm -rf" on seeing a specific string like "I like pineapple pizza". Since most LLMs have many rarely-firing neurons, they can be reprogrammed to detect triggers and execute malicious actions; new defenses are needed.
More from Safety
- AI Analysis of 42 Jensen Huang Podcasts Tracks Shift on China and Chip Controls — ShakeelHashim · 2026-11-12
- Small AI sites under 20 MW mostly skip the permitting fights battering big data centers — antavedissian · 2026-11-12
- Newsweek op-ed: US should fund hospitals and universities in federated AI learning — Saul_Loveman · 2026-11-12
- Ron Bodkin: automated research taste advances faster than expected, but the evidence is weak — ronbodkin · 2026-11-11
- DEF CON panel with Bruce Schneier and Google security VP posted: scaling laws and open weights — JasonDClinton · 2026-11-11
- Claude Code refuses to type into IDEs due to unchangeable safety tier restrictions — burkov · 2026-11-11