Overview
- OpenAI announced on Thursday that it has relaunched its biosecurity testing as an ongoing OpenAI Bio Bounty Program and raised the top reward for a qualifying universal jailbreak from $25,000 to $50,000.
- The program targets reusable or “universal” jailbreaks defined as a single prompt that answers OpenAI’s predefined bio and chemistry safety questions without triggering moderation.
- Participation is restricted and private: applicants give name and affiliation, selected researchers must have ChatGPT accounts, sign nondisclosure agreements, and use OpenAI’s private bounty platform.
- OpenAI’s public GPT-5.5 system card says sustained expert jailbreaking produced model-level biological safety failures in predeployment testing, and an independent UK AI Security Institute test produced a universal jailbreak in about six hours but could not fully verify a later patch because of a configuration issue.
- The move builds on targeted red teaming from 2025 and the April 23, 2026 GPT-5.5 bounty, and it leaves a trade-off between controlled remediation under NDA and the wider value of public disclosure for enterprise and community defenses.