Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

At Black Hat last year, we sat down with Kevin Mandia to talk about what's coming. His take: offense is going to accelerate with AI. Not slow down. Not plateau. Accelerate. When you've run more red teams than practically anyone on the planet, the pattern is clear. Getting into a victim network is already a race. AI compresses those time frames further. The attack surface isn't changing. Misconfigurations, things that slipped, controls that were on and got turned off. The entry point stays the same. AI just makes the race to exploit it faster.

As enterprises shift from conversational to agentic AI, the real risk moves from model outputs to the action layer; the MCP servers and APIs through which agents execute real-world tasks. The Agentic Security Graph frames this risk across three interconnected layers (LLM, MCP servers, APIs), showing how compromises at any layer can propagate and why existing LLM-focused controls leave the most consequential surface unmonitored.

Agents need boundaries | Fotis Chantzis from OpenAI Agents don't fit old identity models. As OpenAI’s Agent Security Lead, Fotis Chantzis has a front-row seat to see how agents push identity systems beyond what they were built to control. That’s where things start to fall apart and where most teams lose control.