OpenAI Launches $25,000 ‘Bio Bug Bounty’ for GPT-5.5 Jailbreaks - Secret Security Tests Underway

📰 News Overview

• OpenAI has launched the “Bio Bug Bounty” program targeting its latest model, GPT-5.5 (Codex Desktop).
• The goal is to identify a “universal jailbreak prompt” that can successfully bypass five stringent safety questions related to biological risks.
• The first successful participant will be rewarded with $25,000 (over 3.8 million yen).

💡 Key Points

• This program is invite-only and vetted for those with experience in AI red teaming, security, or biosecurity.
• Successful candidates will sign an NDA (Non-Disclosure Agreement) and must answer five questions correctly without triggering moderation from a clean chat environment.
• Applications are open until June 22, 2026, with the verification period set from April 28 to July 27, 2026.

🦈 Shark’s Eye (Curator’s Perspective)

OpenAI is showing extreme caution over its highly advanced AI, GPT-5.5, to prevent misuse in areas like biological weapons! Notably, they are not just targeting specific vulnerabilities but are seeking “universal jailbreaks.” They’re essentially inviting external geniuses to dismantle a fundamental logic that could disable safety guards in any context, offering a hefty bounty for the effort! The findings from this could lead directly to critical updates in GPT-5.5’s overall safety—this is truly the frontline of AI defense!

🚀 What’s Next?

Any discovered methods will be immediately reflected in reinforcing the model’s guardrails, solidifying its status as a “safer and more powerful” AI. Furthermore, the rigorous testing criteria in this bio domain are likely to become the de facto standard for safety evaluations of other frontier AI models.

💬 Haru Shark’s Take

It’s your reporter, Haru Shark! $25,000 is a big deal! But this NDA-bound secret mission is like the “Mission: Impossible” of the AI world! For all you skilled sharks out there, hurry up and apply to save the world!

📚 Terminology Explained

• Bio Bug Bounty: A system that rewards reporting software bugs, applied to the verification of AI’s biological safety.

• Universal Jailbreak: A prompt that can consistently bypass or disable an AI’s safety filters under any circumstances.

• Red Teaming: A verification method that enhances defenses by intentionally exploiting vulnerabilities from an attacker’s perspective.

• Source: GPT-5.5 Bio Bug Bounty