Introducing the OpenAI Safety Bug Bounty program

Source: OpenAI Date: 2026-03-25 URL: https://openai.com/index/safety-bug-bounty

Summary

OpenAI’s March 2026 launch of a Safety Bug Bounty program — extending the standard security bug bounty model to AI safety issues specifically. Where conventional bug bounties reward finding security vulnerabilities (authentication bypasses, data leaks, injection flaws), the Safety Bug Bounty rewarded researchers who identified failures in OpenAI’s safety systems: harmful output generation, policy circumvention, jailbreaks that produced dangerous content, and misuse that the safety systems failed to prevent. The program created a financial incentive for external researchers to stress-test safety, not just security.

Implications

Crowdsourcing safety red-teaming at scale. OpenAI’s internal red-teaming was necessarily limited by headcount and adversarial creativity. A public bug bounty program with financial rewards opened safety evaluation to the global security research community — a population with significant motivation and creativity for finding exploits. This scaled the external safety testing program announced in November 2025.

Thread: Safety infrastructure and responsible deployment. Sits alongside the external safety testing expansion, the gpt-oss-safeguard technical report, the Preparedness Framework, and the April 2026 GPT-5.5 bio bug bounty as OpenAI’s systematic effort to build external safety accountability mechanisms.

Watch: What categories of safety failures were in scope for the bounty program, what the reward structure was, and whether the program produced a significant volume of actionable findings that led to model or policy changes.