Lenny's Podcast · December 21, 2025

Why securing AI is harder than anyone expected and guardrails are failing |

with Sander Schulhoff 2.0

Discover why AI security expert Sander Schulhoff warns that current AI guardrails are fundamentally broken and facing an industry correction within 6-12 months. Learn the counterintuitive truth about why humans consistently outperform automated systems at breaking AI defenses, and why traditional cybersecurity approaches fail with AI systems. This episode reveals practical frameworks like CAMEL for AI security, explains when AI vulnerabilities actually matter for your business, and shows how to design AI permissions that account for inevitable attacks. Essential listening for anyone deploying AI systems who wants to understand the real security landscape beyond vendor marketing claims.

Featured insight

You can patch a bug, but you can't patch a brain. If you find a software bug and patch it, you can be 99.99% sure it's solved. Try to do that in your AI system - you can be 99.99% sure the problem is still there. This fundamental difference makes AI security completely different from classical cybersecurity. — Sander Schulhoff

Best for: AI product managers deploying customer-facing systems, Security engineers transitioning to AI security, CTOs evaluating AI security vendor claims

Loading the full episode…