OpenAI Safety Fellowship applications close May 3 — Sep 14 2026 to Feb 5 2027 program structured around scalable oversight and adversarial robustness
OpenAI's Safety Fellowship program closed its first 2026 application cycle on May 3, with the program itself running September 14, 2026 to February 5, 2027. The fellowship is OpenAI's most explicit institutional commitment to external safety research since 2023 — and it lands in the same week as Anthropic's expanded Fellows program, signaling a labs-wide pipeline competition for the next generation of alignment researchers.
The program structure is the substantive piece. Fellows work on scalable oversight, adversarial robustness, and the methodology problems the 2026 International AI Safety Report flagged as central — including evaluation-vs-deployment gap measurement and chain-of-thought faithfulness. The Sep-Feb timing puts cohort outputs into the lab's training pipeline just before the GPT-6 series enters final pre-deployment evaluation. That's the integration loop the Fellows model is designed to optimize: external researchers do the methodology work, the lab internalizes the methods into deployment pipelines, and the next model generation ships with the strengthened evaluation regime.
The competitive context against Anthropic Fellows is what makes this consequential. OpenAI and Anthropic now both run structured external-researcher pipelines with overlapping but distinct focus areas — OpenAI emphasizes scalable oversight and adversarial robustness, Anthropic adds model welfare and AI control. The combined effect is the institutional scaffolding the alignment community has been building toward since 2022: multiple frontier labs with explicit pipelines from academia into deployed safety practice. Expect the cohort outputs from both programs to dominate the methodology literature through 2027.
OpenAI — Introducing OpenAI Safety Fellowship → · Anthropic Alignment — Fellows Program 2026 May and July cohorts → · Alignment Forum — AGI safety research 2025 review and 2026 plans →