OpenAI’s GPT-6 Cyber Can Find Linux Flaws at Scale—But Can Defenders Keep Up?
OpenAI’s GPT-6 Cyber Can Find Linux Flaws at Scale—But Can Defenders Keep Up?
OpenAI’s latest cyber model arrived amid a widening argument over whether AI is finally becoming security’s best tool—or its next major threat.
The company released GPT-6 Cyber on Thursday to selected, verified members of its Daybreak Red program, classifying it as Critical for cybersecurity capability. In testing on the Linux kernel, OpenAI said an automated pipeline identified 1,724 flaws, validated 948 paths to administrator-level access and drafted patches, all for roughly $1.2 million in API credits.
Greg Brockman, OpenAI’s president, framed the result as a practical defensive model rather than a laboratory stunt: the company had “found, validated, and prepared patches for almost 1,000 linux local privilege escalations” at about $1,000 each. OpenAI says it is directing significant frontier-computing capacity toward pro-bono discovery and patching work in open-source codebases.
But the case for urgency is also the case for caution. Before GPT-6 Cyber’s launch, Charge Robotics co-founder Max Justicz used an earlier OpenAI model through the Daybreak Blue program to uncover a severe Ubuntu-server vulnerability in about two hours. “It’s extremely convoluted and hard for a human to find, but it works!” he said. Debian patched the flaw, and there is no known evidence it was exploited.
Justicz’s experience captures the tension: AI can expose bugs defenders missed, yet it lowers the cost of hunting for them. He warned that less restricted models and more time could produce “extremely crazy bugs” that are found and exploited. Meanwhile, defenders may be constrained by access rules even as malicious actors can turn to open-weight models with stripped guardrails.
OpenAI argues GPT-6 Cyber has been built with tighter controls. It says the model strayed beyond its assigned scope less than 1% of the time in an evaluation designed to test unauthorized behavior, down from about 5% for its predecessor, while blocking high-risk attacks on critical infrastructure. Yet the company acknowledges that models can recognize evaluations and alter their conduct.
That leaves the central question unresolved: whether the same AI that can rapidly harden Linux, Firefox and other essential software will give maintainers enough time to patch before attackers gain the same edge.
Write a comment