Sophos

OpenAI's GPT-5.5-Cyber Outperforms Anthropic's Mythos 5

OpenAI updated its security-focused model GPT-5.5-Cyber and published benchmark results showing it outperformed Anthropic’s Mythos 5 on tests designed to find vulnerabilities in large codebases. On the Cybergym benchmark GPT-5.5-Cyber scored 85.6% (vs. Mythos 5 at 83.8% and the prior GPT-5.5 at 81.9%). In an Exploitgym test the updated model reached 39.5% versus 25.95% for GPT-5.5. GPT-5.5-Cyber is part of OpenAI’s Daybreak security initiative and will be made available to registered partner firms (about 30 US companies including Cloudflare, Sophos and Palantir, plus partners in Europe and the UK). OpenAI and Trail of Bits also launched a Patch the Planet review service to audit and fix vulnerabilities in open-source projects.

OpenAI Expands Daybreak Cybersecurity Platform

OpenAI announced a major expansion of Daybreak to accelerate vulnerability discovery and automated patching. Key launches include an updated Codex Security plugin for large-scale scanning, the full release of GPT‑5.5‑Cyber to verified defenders, the Daybreak Cyber Partner Program, and the Patch the Planet initiative (founded with Trail of Bits) to help open-source projects move from findings to fixes. OpenAI says Codex Security has scanned over 30 million commits across more than 30,000 codebases and that human reviewers and automated systems have marked hundreds of thousands of findings fixed. GPT‑5.5‑Cyber achieves new benchmark results on CyberGym, ExploitGym, and SEC-bench Pro. The program includes trusted-access controls and partnerships with security vendors, research groups, and several national governments and EU institutions to enable responsible defensive use and coordinated disclosures.