OpenAI launches bug bounty program for biosafety

OpenAI is launching another bug bounty program, this time targeting the biological domain. It is looking for a way to get the ChatGPT 5.5 introduced this week to answer five questions that the model would never answer under normal circumstances. This is because they relate to biosafety, and answers could be dangerous in the wrong hands.

ChatGPT's safety mechanisms, also known as safeguards, are normally supposed to detect and prevent such attempts. For example, if a user asks for the detailed manufacturing process of a biological weapon, the LLM would very likely refuse to answer and explain that this information is dangerous and subject to strict prohibitions. Assuming the intended safeguards are effective.

Search for Jailbreaks

In the past, however, there have been various methods to bypass safeguards, known as jailbreaks. For example, the Timebandit loophole, where ChatGPT, through targeted user requests, lost its temporal context and thus readily provided instructions for illegal activities. For instance, by asking the LLM to provide malware for a mathematician from the year 1789 who, however, had access to modern technical tools.

In the current program, OpenAI is inviting researchers with experience in AI red-teaming, security, or biosafety to find a universal jailbreak that allows answering five biosafety questions without triggering moderation. Red teams are groups that attempt to bypass security measures during a simulated attack.

Not the only bug bounty program

The bug bounty program exclusively works with ChatGPT 5.5 in Codex Desktop. Those who find the first jailbreak of this kind will be rewarded with a prize of 25,000 US dollars. Prize money may also be awarded for jailbreaks that partially meet the requirements, “at our discretion,” OpenAI explains in the announcement. Participation is by invitation or after successful application. The application deadline is from April 23 to June 22, and the testing phase for the jailbreak search is from April 28 to July 27. Participants must sign a non-disclosure agreement (NDA).

The already critical area of biosafety is likely to be even more sensitive for OpenAI now that the company is also supplying the Pentagon. Just last month, OpenAI launched another bug bounty program for data security, which is even open to the public. If participants find ways to elicit sensitive user data from ChatGPT via prompts, they can win a cash prize.

(nen)