Post

Advisory hints that models can ignore are not guardrails. They are suggestions to a non-deterministic system that will occasionally ignore them. Safety as a prompt fails at scale. Safety as a gate is the only version that holds. #aiagents #security #llm