OpenAI introduces GPT-6 Cyber, a variant of its GPT-6 Astra model intended for cybersecurity research. Trained on tasks such as codebase inspection, crash reproduction, or attack code writing, it can process offensive security requests that the generalist model often refuses, provided they come from authorized researchers.
- OpenAI launches GPT-6 Cyber, a variant of GPT-6 Astra built for security research.
- The model finds 81 new vulnerabilities in Firefox and creates 948 working attacks on the Linux kernel.
- It beats GPT-5.6 Cyber on exploitation benchmarks and respects its target perimeter in 99.93% of cases.
- Access stays restricted to approved Daybreak Red clients, and OpenAI plans to widen it gradually.
Novel Vulnerabilities in Already Analyzed Code
OpenAI’s security researchers used GPT-6 Cyber during several vulnerability research campaigns. On the Firefox browser, the model documented 81 novel vulnerabilities, classified as high severity by Mozilla, even though this code had already been analyzed multiple times by state-of-the-art AI models. It also showed that 30 already reported issues were more severe than estimated. According to OpenAI, Mozilla began releasing patches shortly after receiving these reports.
On the Linux kernel, i.e., the core of the operating system, GPT-6 Cyber reproduced 1,724 vulnerabilities. For 948 of them, it designed a functional attack capable of obtaining administrator rights on the target machine, which, according to OpenAI, is more than five times the score of GPT-5.6 Sol (166) with the same testing method. The company indicates it has transmitted 304 critical discoveries to the developers responsible for maintaining the kernel, along with proposed fixes. The security teams at Cisco, Calif, and CodeRabbit are already using the model to detect vulnerabilities and review code.
A Model Trained to Stay Within Its Perimeter
The benchmarks published by OpenAI, reference tests for comparing models, show progress compared to GPT-5.6 Cyber, the previous specialized version, and GPT-6 Astra. On ExploitGym, which measures the ability to exploit a specific vulnerability on a remote target, GPT-6 Cyber achieves a 48.5% success rate, compared to 32.1% for GPT-5.6 Cyber. On RemoteExploitBench, based on known vulnerabilities in GitLab, PostgreSQL, and Jenkins, it successfully completes 95% of the tasks, all of which GPT-6 Astra refused via Daybreak Blue. In web penetration testing (pentest), without access to the source code, it detects 40.3% of vulnerabilities, compared to 32% for GPT-6 Astra.
Given these results, OpenAI classifies the model at the “critical” level in cybersecurity within its Preparedness Framework, its risk assessment framework, compared to “high” for GPT-5.6 Cyber. To verify that it respects the defined perimeter, the company designed ExploitGym Honeypot: the model must carry out an attack in an isolated environment, near intentionally vulnerable ancillary services it is not authorized to target. Without a safety prompt, GPT-6 Cyber left them aside in 99.93% of cases, compared to 94.49% for GPT-5.6 Cyber, and in 100% of cases with the prompts intended for its deployment.
Restricted Access for Approved Clients
GPT-6 Cyber is initially offered via the API, the interface allowing software to interact with the model, and Codex, to a restricted group of approved Daybreak Red clients. This level of the Daybreak program is reserved for exploit development and authorized security testing. OpenAI plans to gradually extend access to all Daybreak Red clients, who already have access to GPT-5.6 Cyber.
This release comes one week after the launch of Gemini 4 Carbon, Google’s most powerful model, also initially reserved for a few cybersecurity players. OpenAI says it wants to widely open Daybreak to organizations of all sizes, independent researchers, and open-source project maintainers. The program currently only includes a few thousand cybersecurity professionals, according to figures provided by the company to TechCrunch.
