On May 7, OpenAI announced a cybersecurity-dedicated variant in limited preview, available only to vetted teams within a trusted access programme.
The format repeats almost exactly the design a competitor had adopted a month earlier, when it announced and restricted on the same day a model capable of finding vulnerabilities in software audited for decades.
Convergence between rivals on this point indicates a shared problem with no good solution. The capability that finds flaws is the same one that exploits them, and no benchmark separates the two, so what remains is restricting by who accesses rather than by how they use it.
The limit of that approach is familiar. Vetting an organisation isn't vetting a use, and control usually stops at onboarding, without following what happens afterwards inside the approved company.
There's also the asymmetry that ran through the whole year: vetted access programmes reach whoever uses an identified service and don't reach anyone running open-weights models on their own hardware.