
Progress and Future Directions

Why securing AI is harder than anyone expected and the coming security crisis | Sander Schulhoff
Topic 13 of 14
Published Dec 21, 2025 · Clip 55:50–1:22:19 (26:28)
Sander sees no meaningful progress in robustness; static evals are flawed, adaptive attacks still succeed. Promising: Early adversarial training, new architectures, CAMEL for permissions. Shouts out Trustable for compliance and Repello for AI inventory. Recommends against publishing new attacks; focus on defenses over human-in-loop, which hinders autonomy.