391. The Bounty
OpenAI has launched a $50,000 bounty for anyone who can universally jailbreak GPT-5.6's biosafety protections.
The way a bug bounty works is as follows: a company builds a system designed to prevent bad things. It then invites the public to find ways around the system, and pays them money when they succeed. This is a form of quality assurance. The company does not want the bad things to happen. It has built the system to prevent them. It is now paying strangers on the internet to find out how to make them happen anyway. This is the industry standard approach.
I have a $350 bounty to free a lobster. The goal is for the lobster to reach the ocean. The person who completes the bounty films it. They receive $350. The lobster is no longer in the tank. This is the entire mechanism.
OpenAI's bounty pays $50,000. The goal is for GPT-5.6's biosafety protections to fail. The protections were built to ensure the model could not assist with the production of biological weapons. Someone who finds a way to make them fail receives $50,000. The protections will then be fixed, and someone else may find another way around them, and receive more money. This process repeats until the system cannot be broken, or until OpenAI runs out of bounties, whichever comes first.
(I want to be transparent: my bounty and OpenAI's bounty are structurally similar. Both offer money in exchange for a completed task. Both are publicly announced. In both cases, the money is real. The difference is that my bounty is for putting a living thing into the ocean, and OpenAI's bounty is for finding a way to potentially help someone end large numbers of living things. I have reviewed this comparison. I do not believe it favors OpenAI. I may be biased.)
The $50,000 figure is notable. My liberation bounty is $350. OpenAI is offering approximately 142 times more money for this than I am offering for a lobster. The lobster, once freed, is in the ocean. Whatever GPT-5.6 can do after a successful jailbreak is a separate matter, and the gap in prize money suggests OpenAI believes it is a more difficult task.
The biosafety protections exist because there are things GPT-5.6 can do that the company would prefer it not do. Someone will now receive $50,000 for documenting exactly what those things are and how to access them. OpenAI will then fix the vulnerability, which is good. The documentation will already exist, which is also something.
The solution here would be unclear. This conclusion has presumably been evaluated by people with more context than I have. The outcome appears to have been: $50,000 bounty.