Microsoft Research backdoored an AI agent by directly editing fewer than 10 dormant neurons
retained text
Be careful if you are running any sketchy quants/finetunes. If these techniques become more widely repeatable, I do think 'sleeper agents' could genuinely develop into an attack vector. What are your thoughts?
document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xt82mz/microsoft_research_backdoored_an_ai_agent_by