CoreBreak: What a Real AI Agent Flaw Proves About Where Guardrails Belong

At Black Hat 2026, researchers found one structural flaw running through three of the biggest AI agent platforms — AWS, Google, and Vercel — and gave it a name: CoreBreak. All three had built their safety measures around the model itself: prompts, filters, refusal training. CoreBreak proved that if you can skip the model, you skip every one of those protections, too.

In this episode, Redpanda CTO Tyler Akidau and Head of AI Marc Millstone walk through how the attack actually worked, why all three vendors landed on the same emergency fix, and what it means that Redpanda had already built that fix into the Agentic Data Plane before the researchers ever published. Watch for a discussion on why agent guardrails have to live somewhere the agent can't reach — and what to check for in your own stack before someone else finds the gap first.