
Problems with AI Guardrails

Why securing AI is harder than anyone expected and the coming security crisis | Sander Schulhoff
Topic 3 of 14
Published Dec 21, 2025 · Clip 7:32–8:39 (1:07)
Sander describes studying defenses like AI guardrails, which use LLMs to classify inputs/outputs as malicious but fail spectacularly in competitions. He notes guardrails are proposed against prompt injection and jailbreaking but are insecure and ineffective. The segment sets up why traditional software patching doesn't apply to AI 'brains.'