OpenAI's GPT-5.5-Cyber model is launched, outperforming the banned Anthropic Mythos AI amid national security concerns.
OpenAI's New Cybersecurity Model
OpenAI has introduced GPT-5.5-Cyber, a model designed to enhance the identification and resolution of software vulnerabilities. This development comes at a time when Anthropic's AI model, Mythos 5, has been banned by the U.S. government due to national security concerns.
Officially announced on June 22, GPT-5.5-Cyber is part of OpenAI's Daybreak initiative, which aims to bolster cyber defense capabilities. In benchmark testing conducted on CyberGym—a benchmark featuring 1,507 known software vulnerabilities—GPT-5.5 scored a notable 85.6%. In comparison, Anthropic's Mythos achieved 83.8%, while its widely-used model, Claude Opus 4.7, scored 73.1%.
Government Ban on Anthropic's Models
The ban on Anthropic’s models arose from concerns over potential security risks associated with their advanced capabilities. Following an emergency export control order issued by the Trump administration, Anthropic was compelled to take Mythos 5 and Fable 5 offline. The government cited issues with verifying user nationalities and concerns about jailbreak techniques that could bypass AI safety measures.
Despite a narrow performance gap in benchmarks, the implications of these actions are far-reaching. Just days prior to their ban, Anthropic CEO Dario Amodei had described the potential dangers of advanced AI, urging for stringent safety measures similar to those in aircraft regulation.
OpenAI's Strategic Expansion
In contrast to Anthropic’s regulatory challenges, OpenAI is expanding its cybersecurity influence. The Daybreak initiative has formed partnerships with several countries, including Canada, Germany, and Japan, as well as organizations like the European Union Agency for Cybersecurity.
Additionally, 28 security firms, such as CrowdStrike and Cisco, are collaborating with OpenAI to integrate GPT-5.5 into their offerings for specific customers. OpenAI's Codex Security tool has already scanned over 30 million code commits, detecting more than half a million vulnerabilities since its launch.
Nevertheless, GPT-5.5-Cyber remains accessible only to verified cybersecurity professionals. OpenAI conducted preliminary tests with federal agencies before the model's release, showcasing a more transparent approach compared to Anthropic's handling of Mythos.
As of June 23, Anthropic's Fable 5 and Mythos 5 remain offline, with no timeline for their return provided by the company or the Commerce Department.
