OpenAI just launched a safety fellowship for outside researchers

External AI safety researchers usually face an impossible trade-off: stay independent in academia and starve for compute, or join a frontier lab and hand over your autonomy. OpenAI is now trying to pitch a middle ground.

OpenAI announced the OpenAI Safety Fellowship, a new pilot program designed to sponsor external researchers, engineers, and practitioners working on the safety and alignment of advanced AI systems.

Stipends and compute, but hands off internal models

The fellowship will run from September 14, 2026 to February 5, 2027. Selected fellows can work in person at Constellation in Berkeley or participate fully remotely.

OpenAI is sweetening the deal with a monthly stipend, dedicated compute resources, API credits, and hands-on mentorship from its internal staff.

There is, however, a very clear red line. OpenAI explicitly noted that fellows will receive API credits and resources, but will not have internal system access. In other words, you are still poking at models from the outside.

Priority topics include safety evaluations, ethics, robustness, scalable mitigations, privacy-preserving techniques, agentic oversight, and high-severity misuse domains.

Broad admissions with strict deliverables

OpenAI is not restricting applications to standard machine learning researchers.

The call is open to applicants from computer science, social science, cybersecurity, privacy, human-computer interaction, and related disciplines. OpenAI says it prioritizes execution, technical judgment, and research talent over formal credentials, though professional references are required.

Applications close on May 3, with decisions landing by July 25. Fellows cannot simply show up to network: everyone is expected to deliver a tangible output by the end, such as an academic paper, an evaluation benchmark, or a novel dataset.

Why it matters

If you are an independent researcher or graduate student grinding on alignment, this is a compelling opportunity. You get paid, you get precious compute, and you get direct access to OpenAI researchers without having to sign up as a full-time corporate employee.

Strategically, this is a smart play by OpenAI. The lab has faced criticism over its safety culture and turnover over the past year. Funding external fellows and basing them physically around Berkeley hub Constellation helps rebuild goodwill with the broader safety ecosystem.

That said, keeping fellows locked out of internal systems is a glaring limitation. Meaningful alignment work increasingly requires peeking beneath the hood into model activations, training dynamics, and internal states. We think this fellowship will generate great red-teaming datasets and benchmark papers, but as long as the models remain a black box, genuinely fundamental alignment breakthroughs will be hard to come by.