OpenAI unveils GPT‑5.5‑Cyber and Patch the Planet to bolster open‑source security

OpenAI has released the full version of its GPT‑5.5‑Cyber model, a dedicated AI system for automated vulnerability detection, exploit generation and patch creation. In benchmark tests such as UC Berkeley’s CyberGym, the model reproduced 85.6 % of 1,507 known software flaws, surpassing Anthropic’s Mythos 5 (83.8 %). The model is available only through a limited “Trusted Access for Cyber” program to vetted security teams.

Alongside the model launch, OpenAI expanded its Daybreak cybersecurity initiative with the Patch the Planet program, co‑founded with Trail of Bits and supported by HackerOne and Calif. The effort pairs AI‑assisted research with human review to generate and apply patches for more than 30 high‑profile open‑source projects, including cURL, Python, Go, Sigstore and pyca/cryptography. Participants receive ChatGPT Pro, Codex Security plugin access and API credits to automate their workflow.

OpenAI also announced the Daybreak Cyber Partner Program, adding firms such as Darktrace, Check Point, IBM and 28 other security vendors. The program links GPT‑5.5‑Cyber and Codex Security tools to enterprise products. The rollout follows pre‑deployment testing with U.S. agencies (the Center for AI Standards and Innovation and the Office of the National Cyber Director) and partnership agreements with Australia, Canada, France, Germany, Japan, South Korea and EU cybersecurity bodies.

By contrast, Anthropic’s comparable Mythos model has been withdrawn from public use after a U.S. export‑control directive, highlighting the regulatory advantage of OpenAI’s coordinated approach.