OpenAI’s Cybersecurity AI Outperforms Government-Pulled Anthropic Model
OpenAI’s latest cybersecurity artificial intelligence model has surpassed Anthropic’s Mythos AI, which was recently taken offline by the U.S. government, and remains operational.
On June 22, OpenAI officially launched GPT-5.5-Cyber as part of its Daybreak cyber defense initiative. The updated model achieved an 85.6% score on CyberGym, a benchmark developed at UC Berkeley. This test evaluates AI agents on their ability to identify and reproduce 1,507 known software vulnerabilities across 188 open-source projects within a controlled environment.
Anthropic’s Mythos 5 model scored 83.8% on the same leaderboard, while its more widely accessible model, Claude Opus 4.7, registered a score of 73.1%.
While a difference of less than two percentage points might typically be considered minor, the context surrounding these results is significant. Anthropic’s Mythos 5 and Fable 5 models were pulled offline on June 12 following an emergency export control directive issued by the Trump administration, citing national security concerns.
The government’s action was prompted by the discovery of a “jailbreak”—a method to circumvent an AI model’s inherent safety protocols, akin to finding a master key for a secure system. Anthropic’s inability to reliably verify user nationality at scale led to the decision to disable both models globally.
Part of the challenge for Anthropic was self-imposed. The company had spent months promoting Mythos as one of the most powerful and potentially dangerous AI models ever created. Its own launch documentation warned that its cybersecurity capabilities could pose significant risks without adequate restrictions. Anthropic CEO Dario Amodei had also published an essay on June 10, comparing advanced AI models to aircraft that safety regulators should be able to ground if they fail inspections.
Just days after this comparison, the government effectively grounded Anthropic’s AI models.
This was not the only controversy for Anthropic that week. The company had already faced criticism for a hidden filter within Fable 5 that silently degraded the model’s performance for users suspected of developing competing AI technologies, a practice conducted without user notification. Anthropic subsequently issued an apology and reversed the policy.
As Anthropic engages in discussions with the Commerce Department and pursues its lawsuit against the Trump administration, OpenAI is expanding its influence. The Daybreak program has secured cybersecurity partnerships with Australia, Canada, France, Germany, Japan, South Korea, and European Union institutions, including the European Union Agency for Cybersecurity.
Twenty-eight security firms, including CrowdStrike, Cisco, and Cloudflare, have joined OpenAI’s Cyber Partner Program, integrating GPT-5.5 into their products for vetted clients. According to OpenAI’s blog, its Codex Security tool has analyzed over 30 million code commits across 30,000 codebases since its March launch, identifying and logging more than 500,000 fixed vulnerabilities.
The company is also broadening a partner program to enable security firms to incorporate these capabilities into their own tools and has launched “Patch the Planet,” an initiative aimed at addressing vulnerabilities in widely used open-source projects.
It is important to note that GPT-5.5-Cyber is not intended for general public use. Access is restricted to verified security professionals. OpenAI conducted pre-deployment testing with federal agencies, including the Center for AI Standards and Innovation and the Office of the National Cyber Director, prior to its launch. This restricted-access model mirrors Anthropic’s approach with Mythos, but OpenAI obtained government clearance for its strategy beforehand.
As of June 23, Fable 5 and Mythos 5 remain offline, eleven days into their suspension, with no official timeline for restoration provided by Anthropic or the Commerce Department.
Leave a Reply