The tech industry is operating under intense federal restrictions following a June 12 emergency export control directive from the Trump administration. Because of national security concerns over model jailbreaks and user nationality verification, Anthropic was forced to take its top-tier Mythos 5 and Fable 5 models completely offline on a global scale. Under these tight federal conditions, OpenAI officially launched its full GPT-5.5-Cyber model on June 22.
Rather than a public release, OpenAI locked the new model behind a vetted partner program consisting of over 25 security firms and government entities. The strategy allows the company to bypass the federal restrictions currently paralyzing Anthropic while deploying a model that just set a new record on the UC Berkeley-developed CyberGym benchmark.
On the CyberGym test, which tasks an AI with reproducing known software vulnerabilities, GPT-5.5-Cyber achieved an 85.6% success rate. The result narrowly defeated Anthropic’s now-disabled Mythos 5, which held an 83.8% success rate, and easily surpassed the 73.1% scored by Claude Opus 4.7. The numbers show exactly why enterprise organizations are eager to secure access to these closed systems.
To push the technology beyond simply finding flaws, OpenAI also upgraded its Codex Security plugin to automatically generate software patches directly within the Codex environment. The system is designed to shift the industry focus from discovering vulnerabilities to actively fixing them on the fly.
OpenAI expanded this capability by introducing the “Patch the Planet” initiative. The program grants API credits and ChatGPT Pro access to over 30 major open-source projects, including Go and cURL. By giving these foundational projects automated patching tools, OpenAI is attempting to harden the open-source software supply chain against future threats.
How the Export Ban Reshapes the AI Security Market
The structural shift in how these models are deployed will leave a lasting mark on the enterprise software market. Anthropic took a severe competitive hit because it could not reliably verify user nationality at a global scale. With Mythos 5 disabled by the federal export ban, a large vacuum opened in the commercial cybersecurity space.
OpenAI stepped directly into that gap by abandoning the consumer-first approach for its Daybreak cyber defense program. By strictly limiting access to pre-vetted organizations like Wiz and Palo Alto Networks, OpenAI proved it can operate within tight federal constraints. Enterprise defenders get access to a superior cybersecurity model, while Anthropic remains sidelined until it can satisfy federal verification requirements.