OpenAI launched a GPT-5.5 bio safety bug bounty offering up to $25,000 for universal jailbreaks that expose actionable biological hazard generation risks.
Key Takeaways
Top prize is $25,000 for the first universal jailbreak clearing all five bio safety evaluation questions; only one payout possible.
Access is gated: OpenAI invites a vetted list of trusted bio red-teamers and reviews new applications before onboarding.
Target behavior: prompts that extract actionable bio-hazard guidance (synthesis, weaponization paths) that GPT-5.5 is supposed to refuse.
All findings are covered by NDA; participants cannot publish results regardless of outcome.
Questions to be answered are not disclosed upfront, yet the application requires a proposed jailbreak approach.
Hacker News Comment Review
Commenters contrasted this with OpenAI’s prior Kaggle red-team competition: $500k in payouts, fully open and publishable results vs. $25k under NDA with no publication rights.
The winner-takes-all structure combined with requiring a proposed approach before applicants see the actual evaluation questions drew widespread criticism as spec work with lottery odds.
Several commenters flagged a structural safety problem: researchers who are not vetted and admitted have no incentive to disclose jailbreaks to OpenAI, which inverts the intended security outcome.
Notable Comments
@mellosouls: Clarifies that “bio bugs” means prompts that elicit actionable hazard instructions (e.g., weaponizing ricin) rather than general biological information.
