OpenAI Is Launching a Safety Fellowship — and It's Open to Outside Researchers
OpenAI has a credibility problem on safety. The company was founded with safety as its core mission, but has faced years of criticism — from departing employees, from former board members, and from the broader research community — over whether it treats safety as a genuine priority or a marketing talking point. The new OpenAI Safety Fellowship is, at minimum, a concrete step toward opening up.
Announced on April 6, 2026, the Safety Fellowship is a five-month program running from September 14, 2026 through February 5, 2027. It's aimed at external researchers, engineers, and practitioners who want to do rigorous work on AI safety and alignment — people who are not OpenAI employees and won't have access to OpenAI's internal systems.
What Fellows Will Work On
OpenAI is looking for researchers who want to tackle safety problems that actually matter for existing and future AI systems. The stated priority areas are:
- Safety evaluation
- Ethics and robustness
- Scalable mitigations
- Privacy-preserving safety methods
- Agentic oversight
- High-severity misuse domains
The emphasis on "empirically grounded, technically strong" work that's "relevant to the broader research community" suggests OpenAI wants publishable research, not internal tools. Fellows are expected to produce a substantial output — a paper, benchmark, or dataset — by the end of the program.
How It Works
Fellows will work closely with OpenAI mentors and alongside a peer cohort. Physical workspace is available at the Constellation facility in Berkeley, which hosts multiple AI safety research organizations, though remote participation is also allowed.
The program includes a monthly stipend and compute support in the form of API credits. Fellows will not have access to OpenAI's internal models or data — they'll work with the same API access available to outside developers. Letters of reference are required as part of the application.
Applications are open now at OpenAI's website and close on May 3, 2026. Applicants will be notified by July 25. OpenAI welcomes researchers from computer science, social science, cybersecurity, privacy, HCI, and related fields — prioritizing "research ability, technical judgment, and execution over specific credentials."
What This Signals
The fellowship comes at a moment when OpenAI's relationship with the external safety research community has been strained. Several high-profile departures from the company's safety team in 2024 and 2025 were accompanied by public statements about concerns over how safety was being weighted against commercial priorities. The co-founders of Anthropic, OpenAI's most direct competitor, left specifically over safety disagreements.
Opening a fellowship that brings external researchers into contact with OpenAI's work — even without internal access — does two things. It creates a pipeline of talent that has been oriented toward OpenAI's research agenda. And it gives OpenAI something to point to when asked what it's doing to engage the broader safety community.
Whether that adds up to a meaningful change in how the company operates, or is primarily a reputational move, is something only the fellows themselves will be able to assess from the inside. The fact that the program runs at an external facility rather than at OpenAI's own offices is telling: this is explicitly positioned as independent research that OpenAI is supporting, not research that OpenAI controls.