The most helpful AI agent is the least secure one

The most pleasant agentic AI system is the least secure one.

That's the trap Rob van der Veer, Chief AI Officer at Software Improvement Group, lays out clearly: the agent that never says "permission denied" is the agent everyone loves to use. It always works. So we open up all the privileges just to keep it that way.

He says we should do the contrary.

Rob is the founder of the OWASP AI Exchange and co-editor of the forthcoming European AI security standard supporting the EU AI Act. On The Cybersecurity Defenders Podcast, we spoke with him about building trustworthy AI: AI security, threat modeling, international standards, and the new challenges of agentic AI.