GPT‑5.5‑Cyber hits 85.6% on CyberGym — can it patch the planet
openai launched the full gpt‑5.5‑cyber model (achieving 85.6% on cybergym) and codex security tools to automate vulnerability patching, aiming to shift cyber defense from discovery to remediation.
openai announced the full release of gpt‑5.5‑cyber, a model scoring 85.6% on the cybergym benchmark (up from 81.8%), alongside a new codex security update designed to automate the discovery and patching of vulnerabilities in major infrastructure like the linux kernel and freebsd. through the new 'patch the planet' initiative, openai is partnering with trail of bits, hackerone, and 30 open-source projects including curl, go, and sigstore to automate the 'find, fix, and disclose' workflow for defenders. these tools aim to solve the new cybersecurity bottleneck where ai-powered vulnerability discovery outpaces human remediation capabilities, requiring governance and trusted access to ensure defensive use. this launch underscores a strategic pivot from purely finding flaws to accelerating end-to-end patch automation, democratizing advanced defensive capabilities for organizations overwhelmed by security findings.