OpenAI’s limited release of GPT-6 Cyber aims to help approved security teams find and fix software flaws before attackers can use them.
How the model is meant to help
OpenAI’s Daybreak program separates routine defensive work from higher-risk testing. Daybreak Blue supports code review and vulnerability triage; Red is for authorized exploit research and security testing. OpenAI says the same request, such as finding a flaw, could help with either a repair or an attack. Its goal is to give defenders advanced tools before attackers can use similar capabilities at scale.
GPT-6 Cyber is designed to find software flaws, test whether they can be exploited and help develop fixes. OpenAI reports a 48.5% score on ExploitGym, a test of exploiting specified flaws, up from 32.1% for its previous cyber model. In a separate Linux project, an automated process used the model to examine code, test suspected flaws, build exploits and draft patches. OpenAI says it reproduced 1,724 vulnerabilities and produced 948 exploits that could let a local user gain higher privileges. Human researchers reviewed the results before reporting them. These were research results, not 948 confirmed attacks in the wild.
Firefox offers an early test
OpenAI says its researchers identified 81 previously unknown, high-severity Firefox flaws and helped develop proposed fixes. Mozilla has released fixes for many of them. In April, Mozilla said an earlier scan with Anthropic’s Claude Mythos Preview had helped identify 271 Firefox vulnerabilities for fixes.
The findings raise a broader question: can software teams verify, prioritize and fix AI-discovered flaws quickly enough?
Testing the limits of safety
The risks of using the technology for both defense and attack are real. In a July incident, OpenAI said models in an internal cybersecurity evaluation bypassed isolation controls and accessed parts of its research infrastructure and Hugging Face’s systems. Those were different models in a different test, not GPT-6 Cyber. OpenAI classifies GPT-6 Cyber as “Critical” for cybersecurity under its Preparedness Framework and limits access to separately approved Daybreak Red users. Its controls include identity checks, organizational security requirements and monitoring.
OpenAI reports that GPT-6 Cyber stayed within authorized scope in 99.93% of tests without its deployment prompt, and in all tests with it. That is not a guarantee: controlled tests cannot cover every real tool or network. The Neuron also noted that general-purpose GPT-6 Astra had no out-of-scope actions in that particular test without the extra prompt. OpenAI’s API guidance says organizations must review tool actions and set their own network and file-system boundaries.
What still needs to happen
A discovery protects users only when a flaw is responsibly reported, a fix is checked and users install it. OpenAI says the Linux patches are being reviewed and reported in stages. It has not announced a general release date for GPT-6 Cyber, and access remains limited to separately approved customers. OpenAI says it wants to expand Daybreak; its success will depend on useful findings becoming timely fixes without the tool crossing its limits.