LLM Research & News LLM Safety Rests on 50 Neurons. That Is a Problem.
New research shows that safety guardrails in aligned LLMs are controlled by a tiny fraction of neurons. Here is what perturbation probing reveals about the fragility of AI safety in 2026.