OpenAI Unveils GPT-6 Cyber with Restricted Access and Internal Results

OpenAI unveils GPT-6 Cyber, an AI model trained to detect vulnerabilities, develop exploits, and propose fixes. The company claims unprecedented discoveries in Firefox and thousands of vulnerabilities reproduced in the Linux kernel, while limiting access and publishing safety and performance metrics from its own tests. No independent audit has been published yet.

Restricted Access, Active Countermeasures, and Focus on Critical Infrastructure

GPT-6 Cyber is initially offered via the API and Codex to a selected group of Daybreak Red clients, with plans to expand access to all these clients. GPT-5.6 Cyber is available to all Daybreak Red users, while alpha access to Daybreak Blue for GPT-6 Astra is expected to open up to more organizations and individual security researchers. OpenAI reports that the increase in capabilities has been accompanied by a rise in abuse attempts, with some actors trying to transfer these capabilities into their own systems. In response, the company announces checks aimed at aligning the model's actions with user permissions, targeted monitoring of newly disclosed vulnerabilities, and automated controls through "Private Safety Processing." IP-based access restrictions are applied for Daybreak Red, and usage in isolated environments with human supervision is recommended. OpenAI indicates it is paying special attention to critical infrastructures, such as power plants and water treatment facilities. It highlights a billion-dollar funding program, "Daybreak for Frontline Defenders," and a subgroup "Daybreak Red for Critical Infrastructure" for authorized testing. The company mentions a limited window of opportunity to use top-tier models for security purposes and aims to expand Daybreak to organizations of all sizes, independent researchers, and open-source maintainers.

Claimed Discoveries on Firefox and a $1,213,430 Linux Campaign

OpenAI reports that GPT-6 Cyber has uncovered 81 previously unknown critical vulnerabilities in Firefox and reassessed 30 already identified issues as more severe. Mozilla has accepted these reports and deployed fixes. Regarding the Linux kernel, the company states that 1,724 vulnerabilities have been reproduced, and 948 functional exploits developed, compared to 166 for GPT-5.6 Sol under similar conditions. OpenAI claims to have forwarded 304 critical discoveries to maintainers, accompanied by suggested fixes. This campaign relied on a system of reconnaissance and analysis agents: approximately 5,034 sensitive locations were scanned, representing about 7% of the code, then filtered down to 2,308 candidates before confirming 1,724 vulnerabilities in isolated virtual machines and creating 948 exploits. The computational cost is estimated at $1,213,430, or $1,279 per exploit. For comparison, programs like Pwn2Own and TyphoonPWN have historically paid between $20,000 and $70,000 per comparable discovery. OpenAI adds that the agents employed at least ten attack techniques described as unknown to external kernel experts, combining established patterns with approaches presented as novel.

Safeguards: 99.93% Without Instructions, 100% With Active Protections

OpenAI indicates that it has trained GPT-6 Cyber to apply operational safeguards, notably using a honeypot designed from the Hugging Face incident. The company specifies that, without specific security directives, the model adhered to limits in 99.93% of situations, compared to 94.49% for the previous version. With security instructions activated, GPT-6 Cyber would have remained within limits in 100% of tests. OpenAI also classifies GPT-6 Cyber at the "Critical" level in the cybersecurity category of its Preparedness Framework, like GPT-6 Astra, and asserts compliance with the model's operational limits.

ExploitGym Benchmarks and Internal Validations

On the evaluation side, OpenAI announces a score of 48.5% for GPT-6 Cyber on ExploitGym, compared to 32.1% for GPT-5.6 Cyber. The standard model GPT-6 Astra (Daybreak Blue) achieves 42.4%, while GPT-6 Astra with protections activated scores 0.0%. In a test targeting web applications without access to the source code, GPT-6 Cyber reaches 40.3%, compared to 32.0% for GPT-6 Astra. OpenAI mentions that these benchmark results come from its own evaluations and that no independent validation has yet been made public. Additionally, Epoch AI recently demonstrated that an open-source GLM model could exploit known but unpatched vulnerabilities in Firefox. According to OpenAI, GPT-6 Cyber would be capable of spotting vulnerabilities, generating exploits, and proposing corrective solutions, while claiming better performance than its predecessor, with quantified results regarding the Linux kernel.