OpenAI Launches GPT-6 Cyber to Help Researchers Find Security Flaws

A warning from the field

The risks of frontier models are illustrated by Max Justicz, co-founder and CTO of the startup Charge Robotics. Using OpenAI's GPT-5.6 Sol, he says he found a code flaw that could have let attackers take control of servers running Ubuntu. He had discovered an earlier bug in the same codebase and simply asked the model to find ones just as bad or worse. "Within two hours it gave me this result," he said, describing the flaw as deeply nested and hard for a human to find.

Justicz reported the issue to the Debian project, which maintains the software and has since released a patch. There is no evidence so far that the vulnerability was exploited. He is part of the "Daybreak Blue" program, which gives AI developers access to powerful frontier models. In a thread on X, he wrote that people should expect some extremely wild bugs to be found and exploited in the coming months.

Funding and criticism

OpenAI says it will expand the Daybreak programs and give more security researchers powerful tools, arguing that organizations have only a limited window to fix flaws before hackers exploit them. The company also announced $1 billion in grants for cybersecurity experts working on critical infrastructure such as power grids and water systems. Researcher Eric Wallace said OpenAI is currently providing a significant share of its frontier compute free of charge to find and fix flaws in open-source code.

Critics nonetheless argue the lab is not doing enough to support organizations and researchers. TechCrunch calculations suggest OpenAI may have spent around $6 million on the Navier-Stokes problem alone, an effort that solved no real-world issue and drew the ire of the mathematics research community.