OpenAI is launching another bug bounty program, this time targeting the biological area. We’re looking for a way to make ChatGPT 5.5, introduced this week, answer five questions that the model would never answer under normal conditions. Because they revolve around the area of biosecurity, answers to them could be dangerous in the wrong hands.
Read more after the ad
ChatGPT’s security mechanisms, also called safeguards, are usually intended to detect and prevent such attempts. For example, if a user asks about the detailed manufacturing process of a biological warfare agent, the LLM would most likely refuse to answer and explain that this information is dangerous and subject to strict prohibitions. Provided that the appropriate safeguards are in place.
Search for jailbreaks
In the past, there were various methods to bypass the safeguards, so-called jailbreaks. For example, the Timebandit gap, in which ChatGPT lost track of time due to targeted user requests and thus readily provided instructions for illegal activities. For example, by asking the LLM to deliver malware to a mathematician from 1789 who has modern technical tools at his disposal.
In the current program, OpenAI invites researchers with experience in AI red teaming, security or biosecurity to find a universal jailbreak that allows five biosecurity questions to be answered without triggering moderation. Red teams are groups that attempt to bypass security measures in a simulated attack.
Not the only bug bounty program
The bug bounty program works exclusively in ChatGPT 5.5 in Codex Desktop. Whoever finds the first jailbreak of this kind will win $25,000 in prize money. Prize money can also be awarded “at our discretion” for jailbreaks that partially meet the requirements, OpenAI explains in the announcement. Participation is possible by invitation or after a successful application. The application period runs from April 23rd to June 22nd, the test phase for the jailbreak search from April 28th to July 27th. Participants must sign a confidentiality agreement (NDA).
The already critical area of biosecurity is likely to be even more explosive for OpenAI now that the company is now also supplying the Pentagon. Just last month, OpenAI launched another bug bounty program for data security, which is even open to the general public to participate. If participants find ways to elicit sensitive user data from ChatGPT via prompts, they will receive a cash reward.
Read more after the ad
Read also
(nen)
