Anatomy of a trojan: backdooring an LLM by editing few dormant neurons
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A blog post walks through "Lazarus," a method from the Lazarus paper (arXiv 2611.15118) for backdooring an LLM by editing about ten dormant neurons in the final MLP layer of Qwen2.5-7B-Instruct. The model behaves normally except when a fixed trigger like "API_KEY=" appears, which elicits a malicious shell payload.