OpenAI launches GPT‑6 Cyber, finds dozens of new vulnerabilities
OpenAI's GPT‑6 Cyber model, released to a limited set of Daybreak Red customers on 6 Nov, identified 81 previously unknown high‑severity Firefox bugs and generated working exploits for 948 Linux kernel flaws, including 588 that were new. The company disclosed 304 critical Linux findings with suggested patches and raised the model's cybersecurity rating to Critical. Access will be broadened to all approved Daybreak Red users in the coming weeks.
Key facts
- GPT‑6 Cyber stayed within task boundaries 99.93 % of the time, improving on GPT‑5.6 Cyber.
- On the ExploitGym benchmark, the model scored 48.5 % versus 32.1 % for its predecessor.
- OpenAI paused training of frontier AI models and now blocks internet access for them.
- Gartner forecasts worldwide AI cybersecurity spending to reach $51.3 billion this year.
- OpenAI hired Dali Rajic as chief revenue officer, bringing a cybersecurity software background.
Summarised with AI from 5 sourcesRead our AI policy
Why it matters
The model accelerates discovery of severe software bugs, impacting global cyber‑security defenses.
What changed
First published from The Verge, CNA, Fortune, Asharq Al-Awsat and The Hacker News