Top stories
OpenAI launches Daybreak cybersecurity program and updated GPT-5.5-Cyber
- Category: AI
- Status: confirmed
- Sources: Daybreak announcement, GPT-5.5-Cyber trusted access, discussion, The Register
- Summary: OpenAI announced Daybreak on 2026-06-23, a cybersecurity push bundling an updated GPT-5.5-Cyber model, a Codex Security plugin, a roughly 30-partner Daybreak Cyber Partner Program, and a "Patch the Planet" open-source vulnerability initiative run with Trail of Bits and HackerOne. GPT-5.5-Cyber stays in limited preview behind a Trusted Access for Cyber (TAC) identity gate for vetted defenders; OpenAI says it is trained to be more permissive on security tasks rather than significantly more capable than GPT-5.5, and reports CyberGym 85.6% (from 81.8%), ExploitGym 39.5% (from 25.95%), and SEC-bench Pro 69.8% (from 63.1%). Patch the Planet lists more than 30 committed projects including cURL, Go, Python, Sigstore, and pyca/cryptography, and reported 64 pull requests and 51 issues across 19 projects in its first week.
- Comments: The Register noted OpenAI gates GPT-5.5-Cyber to about 30 vetted partners after having criticized Anthropic for similar gatekeeping, and tied the contrast to the ongoing Mythos access dispute.
- Why it matters: It moves AI-assisted vulnerability finding and fixing into named, widely deployed open-source projects such as curl, Go, and Python while keeping the strongest cyber model access-gated, the same trusted-access posture now contested around Anthropic's Mythos.
- Follow-up: Watch Patch the Planet's accepted-fix rate and maintainer-burden reports, and whether TAC access expands beyond the initial partners.