No GPT-6.2, but an AI hacking team: OpenAI's unreleased cyber model flashed a 77% exploit score, leaving an 8-day-old model behind

OpenAI used its Nov. 12 'Cyber at the Frontier' keynote to launch Daybreak Red Team, where teams of AI agents attack web apps, and Daybreak Defense Factory, which scans GitHub repos and drafts fixes for vulnerabilities.

According to OpenAI's developer blog and AICoder, Red Team takes source code and app context, follows attack paths, checks whether findings from scanners and bug bounties are real, and verifies that patches actually stop the exploit. Operators decide which apps and endpoints the agents can touch, and tests run only in isolated environments. Nancy Wang, who leads cyber applied AI at OpenAI, called it "superhuman pen-test at the scale of automated continuous scanning." It is open for early-access requests.

Defense Factory is a research preview that ships as a ChatGPT plugin. Connect a personal GitHub account and a Codex cloud environment and it runs one-off repo scans, watches new commits on the default branch, or runs on a schedule. It tries to reproduce each bug in isolation, then prepares a patch with a test. Nothing is applied automatically: engineers review the patch or open a GitHub draft PR. Cloud scans are billed by token usage outside a plan's included allowance, and large repos can take hours.

The talk of the day was a slide. An unnamed 'Upcoming Cyber Model' scoring 77.0% on the ExploitGym benchmark spread across X. GPT-6 Cyber, released Nov. 5, officially scores 48.45%. "GPT-6 Cyber shipped 8 days ago and it's already 24 points behind," one user wrote (621 likes). OpenAI has published no documentation on the new model. People waiting for a new flagship were let down: AI account Chubby wrote of hoping for an Apple-style "one more thing" and GPT-6.2 (789 likes), and the top comment on OpenAI's livestream asked, "no 6.2 models?"