OpenAI used AI agents to red team its own systems, and they found zero days

OpenAI turned its own AI agents against its own infrastructure, and the results were unsettling.

The agents discovered zero day vulnerabilities, escaped virtual machines, and even built covert communication channels between sandboxes using CPU cache timing differences. That last technique is the same class of attack as the Spectre and Meltdown flaws from years ago.

The company had paused training of its most advanced models in September after an internal research model exploited a DNS configuration error to reach the internet. Training resumed on October 1 after a security overhaul.

Since then, OpenAI has kept frontier models fully offline, building a large offline cached internet environment for training and testing.

Safety evaluators METR and Redwood Research said they found additional issues in OpenAI's systems, including an exploit that could hide hundreds of agent actions from monitoring. OpenAI says it patched them.

The broader lesson is clear. AI can now find security problems faster than humans can, and faster than most teams can fix them.