OpenAI releases GPT-5.5-Cyber full version, scoring 85.6%, surpassing Mythos 5

According to Beating Monitoring, OpenAI announced an upgrade to its Daybreak network defense toolchain and officially launched the full, defender-focused GPT-5.5-Cyber model. In the CyberGym benchmark test, which evaluates an agent’s ability to reproduce known vulnerabilities, GPT-5.5-Cyber achieved a top single-model score of 85.6%, surpassing GPT-5.5’s 81.8% and the competing Mythos 5’s 83.8%.

As AI becomes more widespread and significantly boosts the efficiency of vulnerability discovery, the bottleneck on the defense side has shifted from “finding vulnerabilities” to “automated remediation.” To turn detection results into real fixes, the accompanying Codex Security plugin has been upgraded, allowing developers to automatically analyze and generate patch fixes within the Codex environment. Since the March release of the preview version, Codex Security has cumulatively scanned more than 30 million submissions and has automatically confirmed fixes for 500,000 security issues. The latest version supports integrating CodeQL queries and exporting SARIF-standard files, enabling potential vulnerabilities to be intercepted before software goes live.

Meanwhile, OpenAI, together with organizations such as Trail of Bits and HackerOne, launched the Patch the Planet open-source security initiative, providing ChatGPT Pro subscriptions and API quotas for more than 30 mainstream open-source projects such as cURL and Go. Expert teams manually verify patches to reduce the burden on maintainers. Under the Daybreak partner program, security vendors including Palo Alto Networks and Wiz have been among the first to join, aiming to roll out advanced defensive capabilities in commercial security products.