OpenAI launches Red Teaming Network

OpenAI launched the Red Teaming Network, converting what had previously been ad hoc, per-release recruitment of outside testers into a standing, contracted group it could call on throughout a model’s development cycle rather than only in the weeks before launch.

Earlier red-teaming had produced some of OpenAI’s most cited findings — external testers found that DALL-E 2 amplified racial and gender stereotypes, and that GPT-4 and ChatGPT could be manipulated with prompts designed to defeat their safety filters. The network formalised that work: applicants needed no prior AI experience, only relevant domain expertise in fields such as linguistics, biometrics, finance, chemistry or healthcare, and OpenAI said it would prioritise geographic and disciplinary diversity among members. Depending on a project’s needs, participation could range from a few hours a year to more substantial, individually negotiated engagements, with members compensated for their time and free to collaborate with each other on findings.

The move reflected a broader shift already visible in the GPT-4 system card six months earlier, where external evaluation had become a formal pre-deployment step rather than an informal check. Applications for the network’s first cohort closed on 1 December 2023, with OpenAI aiming to notify successful applicants by the end of that year. The network did not replace internal safety testing or automated evaluations; it added a recurring channel through which outside specialists could probe models neither OpenAI’s own staff nor its automated benchmarks were positioned to test.