Why Defenses Fail

Why securing AI is harder than anyone expected and the coming security crisis | Sander Schulhoff

Topic 11 of 14

Published Dec 21, 2025 · Clip 28:31–43:43 (15:12)

Sander argues automated red teaming always succeeds on transformers, revealing obvious vulnerabilities and scaring non-experts unnecessarily. Guardrails fail due to infinite attack space; 99% effectiveness leaves infinite threats. Adaptive attacks (humans/RL) break them quickly; vendors fabricate stats and ignore non-English attacks. Frontier labs can't solve it, so vendors won't.