OpenAI Launches Safety Bug Bounty for AI Risks

OpenAI has launched a public Safety Bug Bounty program to catch AI abuse and safety issues that fall outside traditional security flaws. This new initiative works alongside their existing Security Bug Bounty, focusing specifically on risks like agentic hijacking (where attacker text tricks AI agents into harmful actions or data leaks), exposure of proprietary reasoning info, and platform integrity issues like bypassing account restrictions. Researchers can submit reports targeting scenarios like third-party prompt injection, unauthorized actions by agentic products, or leaks of OpenAI’s internal data. Jailbreaks are generally out of scope, but OpenAI occasionally runs private campaigns for high-risk areas like biorisk content. Submissions get triaged by dedicated teams, and rewards are possible for flaws that lead directly to user harm with clear fixes. Why it matters: As AI agents become more autonomous and integrated into daily tasks, this program acknowledges that the biggest threats aren’t just code vulnerabilities but how systems can be manipulated to cause real-world harm through unintended behaviors.