Summary
Fireship’s Code Report looks at OpenAI’s new research, “A Thousand Fewer Paths to Root,” in which a swarm of AI agents searched the latest Linux kernel master branch and produced 948 unique root exploits, each with an accompanying patch. The video also covers the model behind it, GPT-6 Cyber, a security-tuned variant of GPT-6 Astra that is not publicly available.
The breakdown explains local privilege escalation and then walks through the four-stage agent pipeline: a scout that narrows the kernel to about 5,000 reachable audit targets, an analyst that surfaced 2,380 candidate bugs at roughly $316,000 in tokens, a triage agent that reproduced them in isolated VMs, and exploit agents that produced 1,169 working root exploits despite protections like KASLR and SMEP. Total cost came to about $1.2 million, or roughly $1,280 per exploit.
The video compares this with an earlier GPT-5.6 Soul run that found 166 exploits at about $3,100 each, notes the reported 48.5% score on Exploit Gem versus 32.1% for GPT-5.6 Cyber, and discusses what such capable cyber models mean for defenders and attackers in the race to secure software.
📺 Source: Fireship · Published November 09, 2026
🏷️ Format: News Analysis
