Technical writing on AI agent security, multi-agent architecture, and what we've learned building and securing production agent systems.
New to AI agent security? Start with the first article and work your way down.
Everyone's focused on model safety. Agent security is a completely different problem. Agents have tools, permissions, network access, and inter-agent communication. The attack surface is infrastructure, not weights.
In multi-agent systems, every agent implicitly trusts the output of other agents. This is the biggest unaddressed vulnerability. Compromise one agent, and the trust chain carries the attack through the entire system.
A deep dive into CSA's Multi-Agent Environment Security Threat Risk and Oversight framework and how we apply it to production agent architectures.
AI agent security isn't new. It's identity management, least privilege, blast radius containment, and zero trust applied to a new actor type. The fundamentals haven't changed. The implementation has.