OpenAI launched the GPT-5.5 biosafety bounty inviting external researchers to surface model behaviors that could enable biological weapon planning, design, or execution, citing the five critical steps previously defined in collaboration with SecureBio, NIST, RAND, and EBRC (https://openai.com/index/gpt-5-5-bio-bug-bounty/). The program builds directly on the 2023 Preparedness Framework's risk assessment for biological and chemical weapons. Primary source coverage is limited to program mechanics while omitting explicit linkage to o1 model's documented gains on scientific reasoning benchmarks that elevate dual-use risk profiles.

OpenAI's 2024 system card for o1 and the Nelson-Rose bioweapon design process paper both document how iterative reasoning loops can compress timelines for pathogen ideation and genetic manipulation; these patterns are absent from the bounty announcement yet central to why external bounties are now required. Related evaluations in the 2024 RAND report "Artificial Intelligence and Biological Risk" and the SecureBio benchmark suite show frontier models still leak restricted information under targeted jailbreaks, a gap the bounty explicitly seeks to close before broader GPT-5.5 deployment.

Mainstream coverage missed the acceleration dynamic: GPT-4 required months of red-teaming; o1 reduced that window through internal chain-of-thought advances; GPT-5.5's bounty indicates OpenAI now treats the capability-safety delta as continuous rather than episodic. Synthesized primary documents confirm investment in scalable external discovery mechanisms as model agency and multimodal biology tooling increase.