← Back to feed

Anatomy of a trojan: backdooring an LLM by editing few dormant neurons

AI · Security · 2026-11-15 14:31

Researchers introduce the Lazarus method, which backdoors large language models by editing a small number of dormant neurons. This technique enables the execution of malicious code when specific trigger words are detected, allowing for stealthy model manipulation without affecting normal performance.

Source: int8.io


Auto-generated summary for reference only. RSS