OpenAI Restricts Access to GPT-5.5-Cyber

OpenAI CEO Sam Altman announced that the company will begin a phased rollout of GPT-5.5-Cyber—an advanced cybersecurity model—exclusively to "key cyber defense professionals." In doing so, OpenAI has adopted the very same gatekeeping approach that Altman publicly mocked Anthropic for just weeks earlier.

From Critic to Follower 

In an April 21 podcast, Altman accused Anthropic of "fear-based marketing" following the company’s decision to limit access to Claude Mythos Preview—its own advanced cybersecurity model announced on April 7 under Project Glasswing. At the time, Altman argued that Anthropic was exaggerating the model's dangers to justify concentrating powerful AI "in the hands of a small and select elite." Now, OpenAI appears to be following the same playbook. "In the coming days, we are beginning the rollout of GPT-5.5-Cyber—an advanced cybersecurity model—to key cyber defense professionals," Altman posted on X. "We will engage with the entire ecosystem and the government to build a system of trusted access in cybersecurity." The identities of the initial recipients remain undisclosed, and OpenAI has provided very few technical details regarding the model's specific capabilities.

British Assessment Ranks Both Models at the Same Level 

This coincides with an assessment by the UK AI Safety Institute published on April 30. The Institute’s experts recognized GPT-5.5 as "one of the strongest models we have tested in cybersecurity tasks"—and only the second model after Anthropic’s Mythos to successfully complete a simulated 32-stage corporate network attack. The Institute estimated that a human would require approximately 20 hours to complete the same task. On expert-level tasks, GPT-5.5 showed an average success rate of 71.4%, while Mythos Preview scored 68.6%, GPT-5.4 scored 52.4%, and Opus 4.7 scored 48.6%. The Institute interpreted these results as evidence that the growth of cyber capabilities is a general trend among frontier models, rather than an isolated breakthrough.