Have OpenAI been on a crusade of "too dangerous to open source" recently? They've pivoted their more to messaging to "competitive reasons" recently, which I appreciate, because it's honest.

FWIW, Gemma4 31B is already quite a capable cyber/security model; do some post-training with RL gyms on it focused on cyber tasks and harnesses for a week or two, and on the specific domain of security/vuln-finding/pen-testing, you'll end up with an extremely capable frontier-cyber model that's entirely under your control at a shocking 31B.

Because securing your codebase, or securing your company's codebase is critical, and I consider it both an ethical and professional responsibility as a developer. It shouldn't depend on whether a classifier fires or not.