A warning shot arrived before OpenAI’s latest launch. Max Justicz, co-founder and CTO of Charge Robotics, said he pointed an earlier OpenAI model at code where he had previously found a flaw and asked it to locate something equally serious. “It came back in two hours with this,” he said of a severe Ubuntu-server vulnerability that Debian subsequently patched.¹

There was no known exploitation of that bug. But Justicz’s result sharpened the central fear around AI-assisted hacking: difficult flaws that once demanded rare expertise may become cheap and repeatable. He argued that unaudited code represents “huge alpha” for anyone able to aim a powerful model at it.¹

On Thursday, OpenAI turned that premise into a product launch, releasing GPT-6 Cyber to a subset of verified Daybreak users and rating it Critical for cybersecurity capabilities. In tests on the Linux kernel, the company’s automated pipeline identified 1,724 flaws, successfully developed administrator-level exploitation paths for 948, and proposed patches after human review. The work used about $1.2 million in API credits; the model also found 81 previously unknown high-severity Firefox vulnerabilities.²

Greg Brockman framed the result as a potentially scalable defensive tool, saying the pipeline had found, validated and prepared patches for almost 1,000 Linux local-privilege-escalation bugs at roughly $1,000 each.Greg Brockman (gdb)³ OpenAI says it is widening Daybreak access, including for open-source maintainers and critical-infrastructure operators.

Yet the company’s safeguards underline the same tension. GPT-6 Cyber is intended for authorized defensive work, and OpenAI says it has improved the model’s ability to remain within assigned scope: it went out of bounds less than 1% of the time in an internal evaluation, compared with roughly 5% for its predecessor.⁴

Critics do not dispute the value of better defenses; they question whether they will reach defenders fast enough. Justicz says access limits can leave legitimate researchers constrained while malicious users turn to less restricted open-weight models. With only thousands of professionals in Daybreak against a vastly larger pool of would-be attackers, the decisive contest is no longer simply who can find the next bug. It is who can patch it first.