OpenAI recently unveiled GPT-5.5-Cyber, a powerful new iteration of their AI model, which achieved a remarkable score of 85.6% on UC Berkeley’s CyberGym benchmark. This performance notably surpassed that of Mythos 5, which scored 83.8%. The context of this achievement is heightened by the recent emergency export control directive that affected both Mythos 5 and Fable 5, underscoring the competitive landscape in AI technology.
“GPT-5.5-Cyber beats Mythos 5 on CyberGym benchmark amid export ban,” highlighting the model’s capability and the ongoing regulatory challenges in the sector.
On UC Berkeley’s CyberGym benchmark, GPT-5.5-Cyber recorded a score of 85.6%. Mythos 5 achieved a score of 83.8% on the same benchmark. Claude Opus 4.7 scored 73.1% on CyberGym. Among these three reported results, GPT-5.5-Cyber holds the highest numeric score and Claude Opus 4.7 the lowest. The listed percentages are the reported CyberGym scores for each model.
The numerical results provide a direct comparison of the models’ performance on the CyberGym assessment. They reflect the benchmark outcomes reported for each model. These figures are the reported CyberGym outcomes for the models listed above, presented as direct numeric results for comparison.
On June 12, Mythos 5 and Fable 5 were pulled offline after an emergency export control directive was issued, and the removal is recorded on that date in the reporting. The directive cited national security concerns as the stated reason for taking the models offline. Both models are identified in the report as affected by that emergency measure.
Daybreak signed cybersecurity partnerships with Australia, Canada, France, Germany, Japan, South Korea and EU institutions, including the European Union Agency for Cybersecurity. Twenty-eight security firms joined Daybreak’s Cyber Partner Program as reported. The Cyber Partner Program includes firms such as CrowdStrike, Cisco, and Cloudflare.
These items are presented in the reported material without technical mechanisms or direct quotations in this section. The report records the June 12 export control directive and Daybreak’s partnership and program details as reported developments.
Codex Security scanned over 30 million commits across 30,000 codebases since launching in March, as reported. The initiative has logged more than 500,000 fixed vulnerabilities in that period. The report lists these totals without providing additional technical details in this section.
Daybreak launched Patch the Planet to fix vulnerabilities in widely used open-source projects. CyberGym is reported to contain 1,507 known vulnerabilities drawn from 188 open-source projects. These programmatic and vulnerability figures are presented in the reported material alongside other developments.
The reported items document recent scanning and remediation activity affecting open-source software. The material presents the numeric totals and program names without detailed technical mechanisms in this section.
OpenAI enters the market with visible momentum, driven by GPT-5.5-Cyber’s leading performance on industry benchmark tests compared with other reported models, positioning the company as a frontrunner in recent model evaluations. At the same time, the sector faces notable regulatory headwinds after an emergency export control directive removed competing products from service, underscoring the intersection of competitive advancement and national-security-driven restrictions in AI deployment.
