The U.S. has blocked Anthropic’s models, but GPT-5.5-Cyber continues to operate and delivers better results

OpenAI has officially launched GPT-5.5-Cyber—a specialized artificial intelligence model designed to detect and fix software vulnerabilities. The announcement took place on June 22 as part of the Daybreak cybersecurity program. On the CyberGym benchmark, developed at the University of California, Berkeley, the new model scored 85.6%—the best result among all tested systems.

CyberGym tests an AI’s ability to reproduce 1,507 known vulnerabilities from 188 open-source projects in a controlled environment. It was on this platform that GPT-5.5-Cyber outperformed Anthropic’s Mythos 5, which scored 83.8%. Anthropic’s more accessible model—Claude Opus 4.7—scored 73.1%.

A difference of less than two percentage points would not, in itself, be sensational. But context changes everything: Anthropic’s Mythos 5 and Fable 5 were shut down on June 12 after the Donald Trump administration issued an emergency export control order citing threats to national security. The trigger was the discovery of a jailbreak technique—a method for bypassing the model’s built-in security restrictions. Since Anthropic lacked a reliable mechanism for verifying users' citizenship, the company shut down both models for everyone without exception.

Anthropic brought some of these problems upon itself. For months, the company had positioned Mythos as one of the most powerful—and at the same time, most dangerous—AI models, warning in its own documentation about the risks posed by its cyber capabilities. On June 10, CEO Dario Amodei published an essay comparing advanced AI models to airplanes, which safety regulators should have the authority to ground based on audit results. A few days later, the government did just that. Additionally, the company faced criticism over a hidden filter in Fable 5 that, without warning, degraded the quality of responses for users suspected of developing competing AI products. Anthropic was forced to apologize and discontinue this practice.

While Anthropic is in negotiations with the U.S. Department of Commerce and continues its lawsuit against the Trump administration, OpenAI is expanding its market presence. As part of its Daybreak program, the company has entered into cybersecurity partnerships with Australia, Canada, France, Germany, Japan, South Korea, and EU institutions, including the EU Cybersecurity Agency. Twenty-eight companies, including CrowdStrike, Cisco, and Cloudflare, have joined the Cyber Partner Program.

According to the OpenAI blog, since its launch in March, the Codex Security tool has scanned over 30 million commits across 30,000 codebases and recorded more than 500,000 fixed vulnerabilities. The company has also launched the «Patch the Planet» initiative to address vulnerabilities in popular open-source projects.

GPT-5.5-Cyber is not a publicly available model: it is intended exclusively for verified cybersecurity professionals. Prior to the launch, OpenAI conducted testing in collaboration with U.S. federal agencies—the Center for AI Standards and Innovation and the Office of the National Cybersecurity Director. It is precisely this approach—coordinating with the government before entering the market—that distinguishes OpenAI’s strategy from that of Anthropic. As of June 23, Fable 5 and Mythos 5 have remained unavailable for the eleventh consecutive day—with no official date for when service will resume.