OpenAI has launched the OpenAI Safety Fellowship to bolster AI safety research. The program is aimed at independent researchers, engineers, and practitioners who wish to conduct research on the safety and alignment of artificial intelligence. The fellowship will run from September 14, 2026 to February 5, 2027 *.
An Open, Multidisciplinary Program
This fellowship is also open to profiles from the social sciences or cybersecurity, thereby broadening the spectrum of potential candidates.
Several priority areas are identified: safety evaluation, ethics, system robustness, large-scale risk mitigation methods, privacy-preserving safety approaches, oversight of autonomous agents, and high-severity misuse scenarios. OpenAI notes that it is prioritizing empirically grounded, technically solid work that is useful to the broader AI research community.
Substantial Resources for Fellows
The program provides a weekly stipend of $3,850 as well as access to approximately $15,000 per month in compute resources to enable selected candidates to devote themselves fully to their work.
Fellows will also benefit from mentorship by OpenAI internal mentors and collaboration with their peers. A workspace is planned in Berkeley, within Constellation, though remote participation is also allowed.
At the end of the program, OpenAI expects each participant to contribute a tangible output: a research article, a new dataset, or a benchmark useful to the AI research community.
A Bug Bounty Component Dedicated to AI Safety
This fellowship sits within a broader strategy unveiled in late March with the public launch of a Safety Bug Bounty, complementing its existing bug bounty program focused on cybersecurity.
This new program concentrates specifically on identifying AI-related abuse risks and safety concerns within OpenAI products. Submissions that pose significant abuse risks, even when they do not meet the criteria of a traditional security vulnerability, are accepted.
The scenarios covered include notably third-party prompt injections, risks associated with autonomous agents such as Atlas Browser, Codex or Operator, as well as flaws in MCP connectors that could be exploited to cause real harm.
Researchers can be rewarded up to $7,500 for reports detailing reproducible, high-severity issues, accompanied by concrete recommendations. Submissions are hosted on the Bugcrowd platform and managed jointly by OpenAI’s Safety and Security teams.
An Initiative in a Competitive Landscape for AI Safety Talent
The launch of this program comes in a context of intensified competition among AI companies to attract top security researchers.
OpenAI’s fellowship structure aligns with the program offered by Anthropic, which proposes exactly the same weekly remuneration ($3,850) and monthly compute allocation ($15,000). This suggests that these amounts have become industry standards for attracting high-caliber researchers specializing in AI safety.
As a reminder, Anthropic has recently announced the opening of applications for two new cohorts of its own program, starting in May and July 2026. The areas now covered include supervision at scale, adversarial robustness, AI control, and model welfare.
*The applications are open until May 3, 2026. Selected candidates will be notified by July 25, 2026.
