Strengthening our safety ecosystem with external testing

Source: OpenAI Date: 2025-11-19 URL: https://openai.com/index/strengthening-safety-with-external-testing

Summary

OpenAI’s November 2025 announcement of expanded external safety testing programs — formalized red-teaming partnerships with external security researchers, academic institutions, and third-party safety evaluators to stress-test GPT-5.1-Codex-Max and other deployed models before and after launch. The expansion represented a shift from primarily internal red-teaming toward a more structured external ecosystem, with defined participation criteria, responsible disclosure processes, and feedback integration into the deployment decision.

Implications

Institutionalizing external safety testing. Pre-deployment external red-teaming had been used for o1 (2024) and GPT-5 (2025), but typically on an ad hoc or invitation basis. Formalizing this into a program with defined relationships and processes made the external testing infrastructure reproducible and less dependent on individual researcher relationships.

Thread: Safety infrastructure and responsible deployment. Sits alongside the gpt-oss-safeguard release, the sparse circuits interpretability work, the Preparedness Framework update, and the chain-of-thought monitorability research as OpenAI’s expanding safety methodology portfolio in late 2025.

Watch: How the formal external testing program compared to the Safety and Security Committee structures, and whether the expanded ecosystem produced material pre-deployment safety findings that altered model or product decisions.