OpenAI builds a dedicated red team network to stress-test models

Before shipping major models, AI labs usually hire specialists to break their systems on purpose. Historically, OpenAI did this in frantic, last-minute testing sprints before big rollouts. Now, they want to turn that fire drill into a permanent fixture.

OpenAI issued an open call for its new Red Teaming Network, inviting global domain experts to help probe and stress-test its models. Applications for this round run through December 1.

Moving from one-off audits to a standing squad

In the past, OpenAI relied on ad-hoc external squads before marquee debuts, like it did prior to deploying DALL-E 2 and GPT-4. Experts came in, found dangerous capabilities, and left once the product shipped.

The new Red Teaming Network formalizes that pipeline across the entire development lifecycle. Instead of a single evaluation right before launch, network members can be tapped at various stages of model training and product rollout.

OpenAI notes that technical AI literacy is not required. The company says it cares most about deep domain knowledge, varied geographic perspectives, lived experiences, and linguistic diversity to spot real-world blind spots.

Five hours a year, paid, under NDA

The gig is designed to be low-friction. Experts do not have to test every single model, and commitments can be as short as 5 to 10 hours across an entire year.

OpenAI also confirmed that participants will be paid for their time on specific campaigns. Between assignments, members will have access to a shared community to swap testing methods and risk taxonomies.

The catch is confidentiality. All testing happens under strict non-disclosure agreements, and some work will remain confidential indefinitely. While OpenAI frequently highlights selected red-teaming takeaways in public System Cards, members cannot openly publish their findings on their own.

For those who prefer open research, OpenAI pointed to its open-source Evals repository, which includes tests for deception, persuasion, and steganography, alongside its Researcher Access Program.

Why it matters

We think this is a pragmatic move by OpenAI to scale up safety, but it comes with clear tradeoffs.

For outside specialists in law, medicine, or national security, this offers rare, compensated access to unreleased frontier models. The ultra-flexible hour requirement makes it accessible to busy professionals. However, the strict NDA terms mean researchers cannot easily leverage their findings into academic papers or public commentary, which may turn away certain open-science researchers.

For the broader AI ecosystem, AI risks have long outgrown pure software engineering. Preventing bioweapon advice or subtle political manipulation requires real-world specialists, not just machine learning researchers. By building a standing pool of vetted reviewers, OpenAI is setting a new baseline that rivals like Anthropic and Google will have to match, raising the cost and complexity of launching frontier models.

Still, let us be clear about what this is. A paid, NDA-bound network selected by the vendor is essentially an expanded corporate QA program, not an independent regulatory audit. Because OpenAI controls who is called up and what findings see the light of day, it solves internal blind spots, but it does not fully solve public trust.