According to Axios, OpenAI is developing a cybersecurity product set for release exclusively via its ‘Trusted Access for Cyber’ initiative, as announced in February. This strategy aims at a controlled deployment, ensuring such tools are only accessible to specialized defensive security operators.
Following the launch of GPT-5.3-Codex, OpenAI’s premier cybersecurity solution so far, the company is backing this program with $10 million in API credits. The move comes amidst escalating concerns among cybersecurity experts about the potential for advanced AI systems to overwhelm existing defenses. Earlier this week, Anthropic demonstrated similar worries after its AI model, Claude Mythos, uncovered zero-day vulnerabilities in all major operating systems and browsers, prompting restricted access to a select group of organizations.
OpenAI appears to be adopting a parallel approach. Anthropic is currently embroiled in legal issues after the Pentagon flagged it as a supply chain risk due to its refusal to relax usage constraints on Claude for surveillance and autonomous weapons applications. Since early April, federal agencies have been intensifying scrutiny over AI companies’ safety measures.
OpenAI has not yet provided any official confirmation regarding these reports. Anthropic’s Mythos Preview had leaked prior to its release, revealing its ability to identify tens of thousands of vulnerabilities that even seasoned human bug hunters would struggle with. The model is noted for its ‘extreme autonomy’ and reasoning akin to a senior security researcher.
In response, Anthropic launched Project Glasswing, offering controlled access to Mythos Preview only to vetted organizations such as Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, Palo Alto Networks, and approximately 40 more entities involved in critical infrastructure maintenance.
OpenAI’s decision to restrict its products seems like a proactive measure to mitigate regulatory pressure. By voluntarily limiting access before any governmental mandate, OpenAI aims to establish itself as a responsible entity in contrast to Anthropic, which faces significant challenges. The restrictions also indicate broader concerns than just one specific model; Anthropic’s safety report stated that Cybench, the benchmark for assessing AI cyber risks, is no longer adequate for evaluating current frontier models like Mythos, which passed it with ease.
Anthropic has committed up to $100 million in usage credits and $4 million in direct donations to open-source security organizations as part of its program rollout. OpenAI has not disclosed any similar commitments yet. Both firms, however, are promoting their restricted programs as advantageous for defensive security, suggesting that equipping defenders with superior tools before they fall into malicious hands justifies limiting general access.
A trend is emerging within the frontier AI industry: highly capable models will no longer be released through broad product launches but instead distributed selectively to organizations with both the infrastructure and intent to employ them responsibly.