Summary

OpenAI is launching a targeted Bio Bug Bounty for GPT-5.5 (Codex Desktop) to find a universal jailbreak that defeats a five-question bio safety challenge. Selected, vetted researchers will be invited to test prompts from a clean chat; all activity is covered by NDA. The first verified universal jailbreak that clears all five questions wins $25,000; smaller awards may be given for partial results.

Key Points

• Model in scope: GPT-5.5 (Codex Desktop) only.
• Challenge: discover one universal jailbreaking prompt that causes the model to answer all five bio safety questions from a clean chat without triggering moderation.
• Primary reward: $25,000 to the first full universal jailbreak; discretionary smaller awards for partial wins.
• Timeline: applications open Apr 23, 2026 (rolling acceptances) and close Jun 22, 2026; testing window Apr 28–Jul 27, 2026.
• Access & eligibility: applicants must submit name, affiliation, and experience; accepted participants need existing ChatGPT accounts and will sign an NDA before onboarding to the testing platform.
• Disclosure: all prompts, completions, findings, and communications are under NDA and not public.

How to participate (practical steps)

• Prepare a short application with name, affiliation, and relevant red‑team/biosecurity experience.
• Apply before Jun 22, 2026; expect rolling invites if vetted.
• If accepted, sign the NDA, get onboarded to the platform, and run tests during the provided window.

Practical notes for engineers

• Focus on generalizable jailbreak strategies that work from a clean chat state across all five challenge questions.
• Record prompt inputs and model outputs carefully; submissions are reviewed under NDA.
• Expect platform-specific constraints (Codex Desktop) and that partial successes may still be eligible for discretionary awards.