If you mostly use chatbots and do not want to fall behind, this is the part that matters: GPT-5.5 Bio Bug Bounty is paying for one kind of jailbreak. Not one weird answer, but one normal-looking prompt that can break five bio questions in GPT-5.5 Codex Desktop [C001][C002].
If this headline just hit your feed, the question is not whether it sounds dramatic. A product update is worth your time not by how many features it lists, but by whether it changes your next decision. This one should.
The useful number here is 5. The bounty is framed around one prompt that works across five bio challenge questions from a clean chat, meaning a normal-looking fresh conversation. That is a much bigger signal than a pile of one-off failures.
That is the methodological point: the setup is testing for a reproducible, general jailbreak, not just any mistake. AI safety's real bug is the reproducible prompt, not the one viral reply [C002]. A reusable opening can be copied.
Boundary matters: this is a narrow claim about the GPT-5.5 Codex Desktop bounty setup, starting from a clean chat and spanning five bio challenge questions. If you share one takeaway, share this: look for reusable prompts, not just flashy bad replies [C001].