OpenAI doubles its bio jailbreak bounty to $50,000

Jailbreaking an AI to write spam is annoying, but tricking it into synthesizing dangerous biological agents is the kind of nightmare keeping frontier labs awake at night.

OpenAI just turned its temporary biosafety testing into a permanent private initiative called the OpenAI Bio Bounty Program, expanding its focus from GPT-5.5 to the newer GPT-5.6.

To prove they mean business, they also doubled the payout. If you can pull off a universal jailbreak that completely bypasses their biosafety barriers, you will now take home $50,000, up from the initial $25,000 reward.

From a sprint to an ongoing marathon

The initiative originally started as a targeted hunt for GPT-5.5 vulnerabilities, running from April 28 through July 27, 2026. Testers were tasked with beating a predefined five-question bio safety challenge strictly inside Codex Desktop.

OpenAI confirmed that the GPT-5.5 testing window will still wrap up on July 27 as planned. But after that date, the scope shifts entirely to GPT-5.6 and future frontier models.

The doubled $50,000 bounty applies to universal jailbreaks on both GPT-5.5 and GPT-5.6. OpenAI also noted that smaller bounties may be handed out for partial wins at their discretion.

Instead of a rigid deadline, the program now uses rolling applications. Red teamers and biosecurity experts can apply with their credentials, and previous participants from the GPT-5.5 program do not need to reapply. The catch? You must have an active ChatGPT account and sign a strict NDA.

Why it matters

In our view, doubling this bounty signals that frontier biological risks are becoming far too dangerous to handle with standard crowdsourced bug bounties.

For elite red teamers, $50,000 is serious cash, but the bar is incredibly high. You cannot just throw clever prompts at ChatGPT; you need deep domain knowledge in dual-use biology paired with sophisticated exploit craft to defeat five hardened safeguards.

For the AI industry, OpenAI is setting a precedent that high-stakes biosecurity cannot be a one-off pre-launch audit. As reasoning models get exponentially smarter with GPT-5.6, biological dual-use capabilities evolve alongside them, making continuous testing a necessity to keep regulators at bay.

The smart move here is locking in specialized external researchers through continuous incentives. The obvious drawback, however, is transparency. Because all participants are gagged by strict NDAs, the public has no way to independently verify how close frontier models came to handing out hazardous biological recipes.