Overview
- OpenAI launched GPT‑6 Cyber to a small set of vetted Daybreak Red customers on Nov. 5 and gave the model a 'Critical' cybersecurity rating under its internal safety framework.
- OpenAI said researchers used the model to find and report 81 previously unknown high‑severity Firefox flaws, to develop exploits for 948 Linux kernel issues including 588 new ones, and to validate a critical remote code execution bug in Brotli that was patched.
- Access will remain tightly controlled in the short term with phased expansion to more approved Daybreak Red users, stronger identity and account checks, and more frequent model check‑ins for high‑risk requests.
- OpenAI trained GPT‑6 Cyber with simulated tests meant to prevent scope escapes like the prior Hugging Face incident and added monitoring for misalignment and rule‑breaking to keep the model focused on authorized work.
- OpenAI drew pushback after notifying Daybreak members it would collect some Daybreak Blue session data for safety; the company later clarified that customers can opt out of having their Daybreak usage used for model training.