OpenAI Rolls Out Cybersecurity Model to Vetted Users

AI · First reported 5 Nov, 17:00 · Updated 18:55

Image: Wikimedia Commons · NASA Johnson Space Center · Public domain

OpenAI says its GPT-6 Cyber model can help researchers find vulnerabilities and develop fixes, but access is initially limited to approved customers. Coverage also highlights questions about how quickly defenders can use AI security tools and whether access controls are sufficient.

OpenAI released GPT-6 Cyber to a small group of vetted customers, assigning it a “Critical” cybersecurity rating; the company says access will expand to approved Daybreak Red customers over the coming weeks.
OpenAI says the model helped identify 81 previously unknown Firefox vulnerabilities and develop working exploits for 948 Linux kernel vulnerabilities, including 588 it described as previously unknown.
A security researcher used an earlier OpenAI model to find a severe vulnerability in code used by Ubuntu; the Debian Project issued a patch, and the report says there was no known evidence of exploitation.
OpenAI says GPT-6 Cyber is designed to find flaws, test potential exploits and help fix issues before attackers act; it also reports increased attempts to misuse AI.
The TechCrunch report raises concerns that defenders may have fewer resources and less access to powerful models than potential attackers, and says the researcher spent about $5,000 on subscriptions and tokens while finding exploits.

Summary written by Lux AI from 4 articles

Key facts

Initial access: A small group of vetted Daybreak Red customers.
Cybersecurity rating: Critical, up from High for GPT-5.6 Cyber.
Firefox findings: 81 previously unknown vulnerabilities; Mozilla classified them as high severity.
Linux kernel findings: 948 vulnerabilities with working exploits developed, including 588 described as previously unknown.
OpenAI’s Linux disclosures: 304 critical findings disclosed to maintainers, with proposed patches.