OpenAI has announced a new OpenAI Safety Fellowship, a pilot program that aims to bring a fresh wave of independent researchers into one of the most contentious and important questions in tech right now: how to keep increasingly powerful AI systems safe, aligned and accountable. It is pitched less like a traditional internship and more like a six‑month, high‑intensity research sabbatical for people who want to work on real safety problems around today’s and tomorrow’s advanced AI models.

The fellowship will run from September 14, 2026, to February 5, 2027, giving fellows roughly five months to design, execute and ship substantial work such as papers, benchmarks or datasets. OpenAI says it is targeting “external researchers, engineers, and practitioners” rather than just students or internal staff, signaling that it wants to widen the safety conversation beyond the walls of frontier labs. In other words, this is an attempt to plant serious safety talent in the broader ecosystem at a moment when both excitement and anxiety over AI’s direction are peaking.

The research agenda is deliberately broad but firmly focused on real‑world issues that current and near‑future systems raise. Priority areas include safety evaluation, robustness, scalable mitigations, ethics, privacy‑preserving safety methods, agentic oversight, and high‑severity misuse domains, among others—essentially, the problems that show up when models become more capable, more autonomous and more deeply embedded in critical workflows. OpenAI is also nudging applicants toward empirically grounded work that can be tested, reproduced and used by the wider research community, rather than purely abstract theorizing.

The structure of the program reflects that ambition. Fellows will be paired with OpenAI mentors and embedded in a cohort, with the option to work in person at Constellation, an independent AI safety and security research hub in Berkeley, California, that already hosts programs like the Astra Fellowship, or to participate remotely. The idea is to give them not just money and compute, but a dense environment of safety‑focused peers, regular seminars and cross‑pollination with other projects tackling similar questions from different angles.

On the support side, OpenAI is offering a monthly stipend, compute resources and ongoing mentorship, plus API credits and other tools where appropriate. Importantly, fellows will not get internal system access, a design choice that keeps the program focused on independent, publishable research rather than proprietary model tinkering. OpenAI stresses that it is prioritizing research ability, technical judgment and execution over specific credentials, and explicitly welcomes applicants from computer science, social science, cybersecurity, privacy, HCI and related fields, with letters of reference required.

The application window is already live and runs until May 3, with successful applicants expected to hear back by July 25. For a company that has been under sharp scrutiny for its internal safety decisions, the timing is notable: the fellowship sits alongside a recent 7.5 million dollar commitment to The Alignment Project, a global fund for independent AI alignment research created by the UK AI Security Institute, which OpenAI frames as part of a broader push to support safety work outside its own walls. OpenAI is careful to emphasize that its funding in that context does not give it control over project selection, which is meant to reassure critics worried about the subtle capture of independent oversight.

Zoomed out, the Safety Fellowship is also a reputational signal. OpenAI’s rapid product cadence, internal reshuffles and dissolution or reconfiguration of some earlier safety structures have led to public skepticism over whether safety still has real teeth inside the company; launching a highly visible pipeline for independent safety talent is one way of answering those doubts without slowing down deployment. It fits neatly with OpenAI’s own stated view that AI safety is a “collective effort” that no single organization can handle alone, and that diverse, outside alignment research is essential as systems approach superhuman capabilities in more domains.

For potential applicants, the program offers a relatively rare combination: direct mentorship from a top frontier lab, a neutral physical base at Constellation, and the freedom to pursue research that is meant to serve the wider safety community rather than a single product roadmap. For the broader ecosystem, the success or failure of this first cohort will be a useful litmus test of whether industry‑funded fellowships can genuinely broaden and strengthen AI safety, or whether they risk becoming just another branding exercise in a field where the stakes keep rising.