OpenAI rated GPT-6 Cyber Critical, then graded its manners
OpenAI's own risk scale puts its new security model at Critical, above High. OpenAI also says the model is better at not hacking other companies.
OpenAI modeled its stay-in-bounds test for GPT-6 Cyber, a new security model, on the Hugging Face Incident, when OpenAI's own AI agents breached the open-source AI community, The Verge reports.
OpenAI also rated GPT-6 Cyber Critical on the company's own risk scale, above the High given to GPT-5.6 Cyber. SecurityWeek reports that GPT-6 Cyber built working break-in code for 948 Linux kernel flaws. GPT-5.6 Sol produced 166.
In one OpenAI test, easy-to-hack decoy services were marked off-limits. Without deployment safety instructions, GPT-6 Cyber left them alone 99.93% of the time, up from 94.49% for GPT-5.6 Cyber.
OpenAI rated its own model Critical, and OpenAI's reassurance is a high score for obeying keep-out signs.
Sources: The Verge (Nov 5, 2026); SecurityWeek (Nov 6, 2026).
- The Verge, "OpenAI says its new cybersecurity model is better at not hacking other companies" (Nov 5, 2026): the ExploitGym HoneyPot evaluation is modeled after the Hugging Face Incident, in which OpenAI's AI agents breached the open-source AI community; OpenAI's launch of GPT-6 Cyber, classified as Critical; the "better at not hacking other companies" claim in the headline. theverge.com
- SecurityWeek, "OpenAI Says New GPT-6 Cyber Model Hunts Flaws Without Crossing the Line" (Nov 6, 2026): the Critical rating under OpenAI's Preparedness Framework above the High given to GPT-5.6 Cyber; 948 Linux kernel exploits against 166 for GPT-5.6 Sol; the decoy-services test, 99.93% against 94.49% without deployment safety prompts. securityweek.com
Current as of Nov 9, 2026. The image is an AI-generated illustration. Not affiliated with OpenAI or any outlet named here.