GPT-6 Cyber: OpenAI Says Its Defender-Only AI Built Root Attacks for 948 Linux Bugs

GPT-6 Cyber is OpenAI’s security AI, open only to approved defenders. OpenAI says it built working attacks on 948 Linux bugs that give a normal user full control, at $1,279 per bug.
cyber-security
artificial-intelligence
Author

Kabui, Charles

Published

2026-11-09

Keywords

gpt-6-cyber, openai, linux-kernel, privilege-escalation, ai-security-research

Download as PDF

GPT-6 Cyber is OpenAI’s security AI for vetted defenders: teams approved for its Daybreak Red program. OpenAI says it used the model on the Linux kernel, the operating system’s core, to build 1,169 working exploits for 948 distinct bugs. An exploit is a program that turns a bug into a break-in. These bugs let an ordinary user become root, the all-powerful administrator. The run cost $1,213,430 in usage fees, or $1,279 per bug by OpenAI’s count.

On OpenAI’s ExploitGym test, which asks a model to turn a named flaw into a working attack, GPT-6 Cyber solved 48.5% of tasks, against 32.1% for GPT-5.6 Cyber. In its trap test, easy-to-hack side services sit beside the target and are off limits. With no safety prompts, it left them alone in 99.93% of cases. In an early run, agents reported 200 root exploits, and human review found 20 used a different unknown bug than assigned. OpenAI says it found 81 new Firefox bugs that Mozilla rates high severity. Mozilla’s Firefox 159 advisory lists 41 fixes credited to OpenAI, without tying them to the 81.

The people who must fix these bugs were already swamped. In May, Linus Torvalds, who leads Linux development, wrote that the flood of AI reports had made the kernel’s security list “almost entirely unmanageable”. No maintainer has commented on this run in the coverage checked. After OpenAI’s 10-day scan, 61 of the root bugs had been patched upstream, 43 had staged patches, and 299 were under public discussion.

Read More: Anthropic’s unreleased model that links bugs into attacks on five major systems and Anthropic’s free bug scanner for open-source projects

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {GPT-6 {Cyber:} {OpenAI} {Says} {Its} {Defender-Only} {AI}
    {Built} {Root} {Attacks} for 948 {Linux} {Bugs}},
  date = {2026-11-09},
  url = {https://toknow.ai/posts/gpt-6-cyber-948-linux-root-bugs/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “GPT-6 Cyber: OpenAI Says Its Defender-Only AI Built Root Attacks for 948 Linux Bugs.” https://toknow.ai/posts/gpt-6-cyber-948-linux-root-bugs/.