OpenAI released GPT-6 Cyber on Thursday, 6 November, a model built for offensive and defensive security work that the company says has already uncovered and validated "thousands" of exploits in critical software. In its launch post, OpenAI said the model found and prepared patches for 948 potential issues in the Linux kernel and separately identified 81 "novel" high-severity vulnerabilities in Firefox, which it says had survived extensive earlier testing with frontier models. On the ExploitGym benchmark, OpenAI reports a 51% improvement over its predecessor, GPT-5.6 Cyber.
Access is deliberately narrow. GPT-6 Cyber is gated behind Daybreak, OpenAI's cybersecurity program, and at launch is limited to a select group of approved "Daybreak Red" customers, the tier meant for penetration testing and advanced security research. A separate "Daybreak Blue" tier covers defensive work. OpenAI says it plans to expand access to all Daybreak Red customers soon and eventually reach smaller organizations, independent researchers and open-source maintainers.
The company paired the release with money for defenders. OpenAI committed $1 billion to support Daybreak participants who protect critical infrastructure such as electrical grids and water systems, its argument that the same capability cuts both ways. As evidence of the cost shift already under way, one researcher reportedly used a prior-generation Daybreak Blue model to find a severe Ubuntu vulnerability, logged as CVE-2026-109518, in about two hours for roughly $5,000 in tokens and subscription fees.
Most of the striking numbers come from OpenAI itself. The 51% benchmark gain and the Linux and Firefox findings are the company's own figures, and "validated thousands of exploits" is a claim that independent maintainers and the affected projects have not publicly confirmed. The 948 Linux "potential issues" are described as prepared patches, not confirmed kernel vulnerabilities, and the scale of that claim will matter only once maintainers accept or reject the fixes. The sharper test is whether the 81 Firefox bugs hold up as genuine, exploitable flaws.
The release sharpens a tension the whole field is circling. A capable vulnerability-finder helps whoever runs it, defender or attacker, and the gating reflects that: OpenAI is handing its most capable offensive model to a vetted few while promising wider access later. The dual-use worry is not hypothetical this week. Separately, researchers at Cisco Talos reported that attackers are using fake event invitations impersonating OpenAI, SpaceX and Anthropic in a phishing campaign to steal Microsoft and Google account credentials, a reminder that AI's brand names are now bait as well as tools.
What remains open is governance, not capability. OpenAI has chosen to control distribution itself rather than wait for any external standard on who should be able to run an automated exploit-finder, and rival labs are weighing similar questions. Anthropic has described comparable internal work as "intermediate progress, and not a completed research agenda," with no current plans to deploy it in production. The gap between that caution and a public launch to paying red-team customers is the story to watch.