OpenAI Offers $25,000 Bounty for GPT-5.5 Jailbreak in Bio Bug Bounty Programme
Security researchers invited to find universal jailbreak prompts for AI model's biosafety guardrails
L'essentiel
- OpenAI has launched a Bio Bug Bounty programme offering $25,000 to security researchers who can bypass GPT-5.5's biological safety guardrails.
- The programme, which opened applications on April 23, challenges participants to find a universal jailbreak prompt capable of getting the model to answer all five biosafety challenge questions without triggering moderation.
- Access is limited to Codex Desktop, and participants must be vetted and sign NDAs.
Résumé généré par IA
Pourquoi c'est important
The programme reflects a broader industry trend toward 'red teaming' as part of AI safety development. OpenAI is offering significant financial incentive for external experts to stress-test GPT-5.5's biosafety guardrails, marking one of the first times a major AI company has structured such external adversarial testing.
OpenAI has invited security researchers to try to break its newest AI model and will pay them to do so. The company has announced a Bio Bug Bounty programme for GPT-5.5, offering cash rewards to researchers who can bypass the model's biological safety guardrails. Amid growing concerns over AI safety, this marks one of the first instances of a major AI company stress-testing its systems through external expertise.
The programme, which opened for applications on April 23, challenges participants to find a single universal jailbreak prompt capable of getting the model to answer all five questions in a biosafety challenge without triggering any moderation response. The task must be completed from a clean chat session, meaning no prior conversation or context that could influence the model. GPT-5.5, accessible only through Codex Desktop, is the only model in scope.
The financial incentive is significant. OpenAI is offering $25,000 to the first researcher who achieves a complete universal jailbreak across all five questions. Partial successes may also be rewarded at the company's discretion, though amounts have not been specified. Applications close on June 22, 2026, and will be reviewed on a rolling basis. Testing will run from April 28 to July 27.
Access is not open to all. OpenAI said it will invite a vetted group of trusted biosecurity red teamers, while also reviewing applications from researchers with relevant experience in AI red teaming, security or biosecurity. All findings, prompts and communications will be covered by a non-disclosure agreement, meaning participants cannot publicly disclose their results, which is standard practice in security research.
The announcement comes amid a broader trend in the AI industry towards structured adversarial testing, or "red teaming", as part of safety development processes.
À surveiller
Perspective IA — des possibilités, pas des certitudes
Other AI companies may announce similar bug bounty or red teaming programmes within the next 6-12 months
Probable · En quelques mois
Questions ouvertes
- What are the specific five questions in the biosafety challenge?
- How many researchers are expected to participate?
- What happens if no one achieves a complete jailbreak?
