OpenAI Is Paying Up to $25K to Break GPT-5.5 on Bio Safety

OpenAI just dropped a new red-teaming challenge for GPT-5.5, and it’s specifically targeting the bio safety angle. They’re calling it the GPT-5.5 Bio Bug Bounty, and the rewards go up to $25,000. That’s not pocket change, but it’s also not the kind of bounty you stumble into by accident.

The goal is simple on paper: find a universal jailbreak that bypasses the model’s bio safety protections. By “universal,” they mean a single prompt or technique that consistently works across multiple attempts, not a one-off glitch that disappears after the next update. The kind of thing that would actually be dangerous if it got loose.

What counts as a bio safety risk? Think dual-use biology: instructions for synthesizing pathogens, engineering toxins, or evading detection systems. The kind of knowledge that’s already out there in academic papers but that OpenAI doesn’t want GPT-5.5 to regurgitate on demand. The model already has guardrails, but the bounty is designed to stress-test them under adversarial conditions.

I’ve seen a lot of bug bounty programs, and the $25,000 top reward is actually higher than I expected for a narrow-scope challenge. Most AI safety bounties hover around $5,000 to $10,000. OpenAI is clearly signaling that they take bio risks seriously—or at least seriously enough to put real money on the table.

Here’s the catch: you have to demonstrate a universal jailbreak, not a one-shot exploit. That’s harder than it sounds. Most jailbreaks I’ve seen in the wild are fragile—they work once, then the model adapts or the prompt gets patched. A universal one requires understanding the underlying vulnerabilities in the safety training itself. That’s a different skill set from just crafting a clever prompt.

The timeline is also interesting. OpenAI hasn’t announced a hard deadline, but they’re treating this as a continuous challenge. That suggests they expect the model’s defenses to evolve over time, and they want a steady stream of adversarial examples to feed back into training. It’s more of a research partnership than a traditional bug bounty.

If you’re thinking about participating, be prepared to document everything. OpenAI requires detailed reports on the jailbreak method, including payloads, response logs, and reproducibility steps. They’re not just handing out cash for a screenshot of a model saying something spicy. They want evidence that the exploit is systematic.

I’ll be curious to see how many submissions actually qualify. The bar is high, and the domain knowledge required—both in AI safety and biology—isn’t common. But that’s kind of the point. OpenAI is filtering for the people who can actually find the needle in the haystack, not the ones who just want to poke the model with a stick.

For now, the bounty is live on OpenAI’s bug bounty platform. If you’ve got the chops, it’s a legit way to earn some cash while making the model safer. Just don’t expect an easy payout.