OpenAI pays up to $25,000 for bio jailbreaks in GPT-5.5, but proof will be in aggregate results

OpenAI launched a bug bounty focused on biological safety risks in GPT-5.5, with rewards up to $25,000. The goal: find universal jailbreaks, not just individual filter bypasses.

The bio bug bounty targets systematic breakdowns, not one-off tricks

The program focuses on universal jailbreaks, meaning approaches that systematically circumvent safety controls across varied inputs, not just a single edge case. Biological safety is one of the most sensitive categories: if a model can provide expert-level biological procedures or help bypass safeguards in specific domains, standard safety evals may not catch it. The bug bounty moves the search outside the internal team.

For the research and security community, this is legitimizing adversarial testing

A formal bounty program signals that OpenAI acknowledges internal red-teaming is not sufficient and that it is willing to pay and listen to external findings. A $25,000 reward for a critical finding is not symbolic. The program also defines scope and disclosure rules, which matters to researchers: they know what they can publish and what they cannot.

The program points in the right direction, but its impact depends on what OpenAI does with the findings

A bounty program is a signal in the right direction, but has limits. Quality depends on what is in scope, how OpenAI handles findings after the deadline, and whether fixed attack classes are incorporated into the next cycle. If the program captures serious findings that OpenAI neither publishes nor incorporates into a public safety report, it becomes a PR exercise. The source page returned 403 during verification.

Aggregate results and adoption by other labs will show the real impact

Watch whether OpenAI publishes results in aggregate form, which attack classes were found, and whether Anthropic, Google DeepMind, or Meta launch similar bio safety programs. Biological safety needs shared data, not a force field owned by one laboratory.

Lilith's verdict

A bio safety bounty is a good step. But impact is measured by what OpenAI does with the findings after the deadline, not by how much it pays for the discovery.