OpenAI has doubled the maximum reward under its specialised Bio Bug Bounty Program to $50,000, as the ChatGPT maker looks to strengthen safeguards around its most advanced artificial intelligence models.
The company said the highest payout will be awarded to researchers who discover a universal biosafety jailbreak - a prompt or technique that consistently bypasses the model's safety guardrails and enables it to generate restricted biology-related information.
The move marks an expansion of a programme that was initially launched as a limited challenge for GPT-5.5. OpenAI said it is now turning it into a permanent initiative, with testing beginning on its latest frontier model, GPT-5.6. The company said the goal is to continuously identify and fix vulnerabilities before they can be exploited in the real world.
Unlike traditional bug bounty programmes that focus on software vulnerabilities, the Bio Bug Bounty is designed to uncover weaknesses in an AI model's safety behaviour. Researchers are asked to find prompts or methods that reliably defeat the model's biological safety protections, helping OpenAI improve its defences against misuse.
Participation in the programme is restricted. Interested researchers must apply, meet OpenAI's eligibility criteria and sign a non-disclosure agreement before gaining access to the testing environment. The company said it may also issue smaller rewards for findings that do not fully qualify as universal jailbreaks but still provide useful insights into improving model safety.
OpenAI has increasingly relied on external researchers to stress-test its AI systems before and after release. The latest increase in the reward reflects the growing importance of AI safety as models become more capable and are used by millions of people worldwide.
The announcement comes amid broader industry efforts to strengthen safeguards around frontier AI systems. As companies race to build more powerful models, governments and independent experts have repeatedly called for rigorous testing to identify potential risks before such systems are deployed at scale.
