OpenAI launches a less restricted GPT-5.4 Cyber model, but you won't find it on ChatGPT

OpenAI has launched its GPT-5.4 Cyber model, which is a specialised version of its GPT-5.4 model built especially for defensive cybersecurity tasks. The new model announcement by OpenAI comes just weeks after rival Anthropic announced its Mythos AI model but did not release it to individual users owing to the risk of misuse.

In a blog post on Tuesday, OpenAI said that it is releasing GPT-5.4 Cyber ‘in preparation for increasingly more capable models from OpenAI over the next few months’. The company said it is fine-tuning its models specifically to enable defensive cybersecurity use cases.

“we aim to make advanced defensive capabilities available to legitimate actors large and small, including those responsible for protecting critical infrastructure, public services, and the digital systems people depend on every day,” the company noted in its blog post.

Since the model is ‘more permissive’, OpenAI is tightly controlling its rollout. GPT-5.4 Cyber will not be available to use via ChatGPT. Instead, OpenAI has begun deploying the model to vetted security vendors, organisations, and researchers.

The rollout of GPT-5.4 Cyber is part of the Trusted Access for Cyber (TAC) programme, a cybersecurity initiative that the company unveiled earlier this year.

Individual users can request access to GPT-5.4 by going to chatgpt.com/cyber and verifying their identity, while enterprise teams must request trusted access through their designated company representatives.

The company noted that access to GPT-5.4 Cyber could come with limitations, especially with no-visibility uses like Zero-Data Retention⁠. This is where organisations access models through third-party platforms and OpenAI lacks direct visibility into the user, the environment, or the purpose of the request.

“Our cybersecurity defenses are the result of many months of iterative improvement. We believe the class of safeguards in use today sufficiently reduce cyber risk enough to support broad deployment of current models,” OpenAI said.

Anthropic had announced its Mythos model last week as part of its ‘Project Glasswing’ and gave access to around 40 organisations like Apple, Google, and Microsoft for defensive cybersecurity purposes.

The AI startup noted that Mythos was able to identify ‘thousands of additional high- and critical-severity vulnerabilities’, including a 16-year-old bug in FFmpeg and a 27-year-old flaw in OpenBSD.