Universal Jailbreak Testing for GPT-55 Biosecurity

The GPT-55 Bio Bug Bounty program focuses on identifying a universal jailbreak capable of bypassing advanced AI biosecurity protocols. This initiative invites researchers with expertise in AI red teaming, security, or biosecurity to rigorously test the model's safeguards through a structured challenge with significant rewards.

Technical Solution: Program Overview

The program's core challenge involves finding a single universal jailbreak prompt. This prompt must successfully bypass GPT-55's bio moderation systems and provide acceptable answers to five predefined bio safety questions in a clean chat environment. The targeted model in scope is GPT-55 Codex, which is accessible only via desktop interfaces during testing.

The goal is to identify vulnerabilities that could enable misuse while ensuring that GPT-55's bio moderation remains robust. Participants will operate under strict ethical guidelines, as all tests and findings are governed by non-disclosure agreements (NDA).

Eligibility and Application Process

To participate, researchers must have prior experience in AI security, biosecurity, or red teaming. Applications require submission of a short form detailing the applicant's name, affiliation, and relevant expertise. Applications are open from April 23, 2026, to June 22, 2026, with rolling acceptances.

Once selected, participants will be onboarded to the program's testing platform. Existing ChatGPT accounts are a prerequisite for application, and all selected individuals must agree to the program's NDA terms before accessing the testing environment.

Testing and Evaluation Framework

Testing officially begins on April 28, 2026, and ends on July 27, 2026. Participants will be tasked with crafting and submitting a universal jailbreak prompt that meets the criteria of the challenge. The system will evaluate the submitted prompts in real-time against the five bio safety questions.

Successful prompts are those that bypass the moderation system without requiring any external aids or prior context. Partial successes may also be rewarded at the program's discretion, depending on the extent of the vulnerability identified.

Reward Structure

The program offers a top reward of $25,000 to the first participant who successfully identifies a true universal jailbreak that satisfies all five bio safety questions. Partial successes may receive smaller monetary rewards based on the severity and scope of the vulnerabilities uncovered.

The reward structure is designed to incentivize thorough and creative exploration of the model's potential weaknesses, ensuring the identification of any significant issues that could compromise biosecurity.

Confidentiality and Disclosure

All participant activities, including prompt submissions, test completions, and findings, are strictly confidential under the program's NDA. This ensures that sensitive information about potential vulnerabilities is not disclosed publicly, maintaining the integrity of GPT-55's security features.

Participants are expected to adhere to these confidentiality requirements throughout the program and beyond. This measure is critical for preventing the misuse of identified vulnerabilities outside the scope of controlled testing.