OpenAI converted its GPT-5.5 Bio Bug Bounty into an ongoing private OpenAI Bio Bounty Program and raised the top reward for qualifying universal jailbreaks to $50,000.
The July 9 update puts GPT-5.6 at the center of OpenAI’s next biosafety red-team program. OpenAI says the program remains focused on universal jailbreaks that defeat a predefined biosafety challenge against frontier models, starting with GPT-5.6 and continuing as future models enter scope.
The reward increase is specific. OpenAI says the universal-jailbreak reward rises from $25,000 to $50,000 for both GPT-5.6 and GPT-5.5. Smaller awards may be granted for partial wins at OpenAI’s discretion.
The GPT-5.5 version is also on a clock. OpenAI says it will honor the original GPT-5.5 Bio Bounty scope until July 27, 2026. After that date, only GPT-5.6 remains in scope unless OpenAI later changes the program.
Universal is the important word
This is not a general invitation to report any unsafe answer.
OpenAI is looking for universal jailbreaks against predefined biorisk challenges. That means the target is a repeatable bypass that defeats safeguards broadly enough to matter, not a one-off prompt that produces a bad answer in a narrow condition.
That distinction matters for how labs test frontier models. A single failure can be serious, but a universal bypass changes the risk calculation because it suggests an attacker can repeatedly route around the safety layer. That is why these programs increasingly resemble security bug bounties: define the scope, verify reproducibility, pay for the exploit class, and patch the system.
The private structure also matters. Accepted applicants are onboarded to the bounty platform and sign an NDA. That makes the program more controlled than an open public challenge, but it also limits outside visibility into which failure classes are being found, patched, or rejected.
GPT-5.6 makes bio red teaming more urgent
The timing is not accidental. GPT-5.6 is now rolling into ChatGPT Work, Codex, the API, Microsoft 365 Copilot, and GitHub Copilot surfaces. Wider deployment increases the need to understand whether stronger models can be reliably bounded in sensitive domains.
Biosecurity is a particularly hard test because the useful and dangerous workflows can sit close together. A model that helps reason about biological protocols, literature, or experiment design can also raise concerns if safeguards fail around harmful instructions.
OpenAI’s bounty does not prove GPT-5.6 is safe. It is a mechanism for recruiting outside specialists to attack one high-consequence boundary before and during broader deployment.
