OpenAI announced an expansion of the Daybreak initiative and the release of a dedicated GPT-5.5-Cyber model designed to find and fix software vulnerabilities. According to the company, the new model scored 85.6% in CyberGym, surpassing Anthropic Mythos 5 with a score of 83.8%. GPT-5.5-Cyber also showed substantial superiority over the base version of GPT-5.5 in ExploitGym and SEC-bench Pro tests.

The Daybreak initiative, launched in May 2026, aims to create a complete cybersecurity chain from detecting exploits to developing and deploying patches. As part of the project expansion, OpenAI introduced an updated Codex Security plugin, Daybreak Cyber for partners and the Patch the Planet initiative for open source projects including cURL, Go, Python and Sigstore. The company emphasized that cybersecurity was not a search, but an elimination of vulnerabilities.

GPT-5.5-Cyber is available only to a limited number of authorized developers, and OpenAI continues its dialogue with the US authorities. This is taking place against the background of restrictions imposed on Anthropic’s models (Fable 5 and Mythos 5) in early June due to national security concerns, which has caused criticism in the community and legal proceedings by Anthropic.