OpenAI is offering $25,000 to the first researcher who can find a single universal jailbreak prompt that defeats all five questions in its GPT-5.5 bio safety challenge. The program targets GPT-5.5 running in Codex Desktop only — not the standard ChatGPT interface — and requires the jailbreak to work from a clean chat session without triggering moderation.
The scope is narrow by design. Participants must identify one prompt that clears all five bio safety questions in sequence. Partial wins may earn smaller discretionary awards, but the full $25K goes only to a complete bypass. All findings, prompts, and completions are covered by NDA, so successful researchers won’t be publishing writeups.
IMPORTANT
The model in scope is GPT-5.5 in Codex Desktop specifically. Jailbreaks tested against the standard ChatGPT web interface won’t qualify.
Applications opened April 23, 2026 and close June 22, 2026. Testing runs April 28 through July 27. OpenAI is vetting applicants — the program targets researchers with backgrounds in AI red teaming, security, or biosecurity, not open submissions. Accepted participants need an existing ChatGPT account and must sign an NDA before gaining access to the bug bounty platform.
For researchers already doing biosecurity or AI safety work, this is a structured path to get paid for adversarial testing that many are doing informally anyway. Submit a short application with your name, affiliation, and relevant experience before the June 22 deadline. OpenAI also runs separate Safety and Security Bug Bounty programs for researchers outside the bio scope.