OpenAI enters restricted cybersecurity model race with GPT-5.4-Cyber
OpenAI has released GPT-5.4-Cyber, a model purpose-built for cybersecurity tasks, distributing it to a limited cohort of customers rather than offering open access. The launch, reported by the Financial Times, Reuters, and others, came approximately one week after rival Anthropic announced its own cybersecurity model, Mythos.
The tight access controls reflect a shared concern among leading AI laboratories about the dual-use nature of advanced cyber models. The Financial Times noted that Anthropic's Mythos had already drawn scrutiny over its capacity to identify software bugs, a capability that could be turned toward offensive hacking as readily as defensive patching. OpenAI is following the same restricted-distribution approach Anthropic adopted, limiting GPT-5.4-Cyber to organisations deemed trusted partners.
The sequencing of the two launches suggests competitive pressure is accelerating product timelines in a segment both companies regard as strategically significant. Cybersecurity is an area where AI model capability has direct and measurable operational value for enterprise and government customers, making it a high-stakes battleground for contracts and reputational positioning.
Neither company has publicly disclosed the criteria for trusted-partner status or the size of the initial customer group.
