OpenAI Finally Tells Us How Human Red Teams Probe Its AI

OpenAI published a white paper and a companion study that pull back the curtain on how the company pressures-test its frontier models before they ship. The white paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” is a kind of playbook for the human side of the process — the part where lawyers, natural scientists, and cybersecurity specialists are given early access and told to break things. It’s a system that’s evolved considerably since the company’s first external red team engagement for DALL·E 2 in early 2022, which was a mostly manual, people-driven affair.

What’s notable here is the level of scaffolding OpenAI now puts around these campaigns. The company starts with a threat-modeling step based on expected model capabilities and previously observed failures, then uses those priorities to decide exactly which external experts to recruit. The paper walks through decisions that sound deeply operational but matter enormously for outcomes: which model version to hand over, whether safety mitigations are already in place, and what kind of interface testers use. An API suits rapid programmatic probing; a ChatGPT-like consumer interface is better for simulating the messiness of real user behavior.

OpenAI emphasizes that no single red teaming exercise can capture every risk — especially risks tied to cultural nuance or regional politics — which is the argument for building reusable evaluations from each campaign. The data these testers generate doesn’t just inform a go/no-go decision for one release. It gets fed back into automated evaluations that can be run repeatedly against future models. That’s the bridge to the companion research study, which introduces a new automated red teaming method. The company’s logic is straightforward: more powerful AI should be able to scale the discovery of model mistakes faster than any group of humans can.

The subtext here is as interesting as the documentation itself. OpenAI made a voluntary commitment last July alongside other leading labs to invest more in red teaming, and publishing methodology is part of that transparency push. But reading between the lines, the white paper also reads like a response to critics who argue that external testing windows are too short or too constrained to be meaningful. By being explicit about how they choose testers, when in the development cycle models are handed over, and how feedback becomes evaluation infrastructure, OpenAI is making the implicit case that their process has grown up considerably.