OpenAI is finalizing a new model with advanced cybersecurity capabilities and plans to release it only to a small set of companies, according to Axios and a source familiar with the matter. The move would mirror Anthropic’s limited rollout of its Mythos Preview model. The decision reflects growing concern that more capable AI systems could be misused for hacking.
The report says OpenAI has already been testing a narrower access approach through its Trusted Access for Cyber pilot program. That program was introduced in February after the rollout of GPT-5.3-Codex, described as the company’s most cyber-capable reasoning model. OpenAI said invite-only participants could use more permissive models to support legitimate defensive work.