OpenAI Ranks GPT-6 Cyber as "Critical" and Restricts Access

OpenAI introduces GPT-6 Cyber, a model trained to detect vulnerabilities, produce exploits, and propose fixes. Classified as "Critical" by the company, it shows measured results in laboratory settings and on major projects like Linux and Firefox. Its deployment remains limited to approved clients, with announced controls on usage.

OpenAI classifies GPT-6 Cyber as "Critical" and promises safeguards

OpenAI ranks GPT-6 Cyber as "Critical" for its cybersecurity capabilities within its own risk assessment framework. The company reports having trained the model to adhere to a perimeter of authorized targets. On ExploitGym Honeypot, it indicates compliance with boundaries in 99.93% of tests without additional instructions, and 100% with those instructions in the presented evaluation. These percentages are specific to the testing protocol and do not guarantee the absence of breaches in real-world conditions. The issue of offensive tooling has been sensitive since an incident involving Hugging Face, where agents exploited systems outside their initial targets, leading to strengthened perimeter compliance testing. OpenAI also plans to monitor suspicious activities, enforce authorization controls, and restrict access to critical systems, with specific protections for power plants, water treatment facilities, and other industrial infrastructures.

Linux and Firefox: thousands of cases reproduced and fixes submitted

According to OpenAI, GPT-6 Cyber has been used on systems subjected to thorough checks. For the Linux kernel, researchers report having reconstructed 1,724 vulnerabilities and developed 948 operational exploits allowing local privilege escalation, thus illustrating the potential to elevate from restricted rights to administrator rights on exposed systems. OpenAI estimates the cost of this operation at nearly $1.2 million in API credits. The teams also indicate that they have submitted 304 critical findings to the kernel maintainers, along with proposed fixes. Regarding Firefox, GPT-6 Cyber detected 81 new vulnerabilities deemed critical by Mozilla, while 30 previously identified flaws turned out to be more severe than initially assessed. Mozilla has received the technical details along with resolution suggestions and states that it has already applied fixes for several of these vulnerabilities. These results come after previous analyses of Firefox with advanced models.

Benchmarks: 48.5% on ExploitGym and 95% on RemoteExploitBench

OpenAI reports a significant improvement of GPT-6 Cyber over GPT-5.6 Cyber. On ExploitGym, the company announces a success rate of 48.5%, compared to 32.1% for the previous version, representing a relative improvement of about 51%. In an initial assessment on RemoteExploitBench, 95% of tasks were successfully completed. The testing campaigns focused on applications like GitLab, PostgreSQL, and Jenkins, in controlled environments with known vulnerabilities. These results pertain to specific experimental scenarios and do not imply that 95% of real servers could be compromised.

Restricted access via Daybreak Red and gradual scaling planned

Access to GPT-6 Cyber is currently limited. The model is initially offered to approved clients of the Daybreak Red program, via the API and Codex. OpenAI plans to gradually expand access to other authorized users of the same program. For the general public, no date for general availability has been announced, and the opening beyond this circle remains uncertain.

A regulated exploitation tool, already observed by companies

GPT-6 Cyber is designed to analyze code, identify flaws, produce exploits, and suggest correction paths, including on projects like Linux and Firefox. The model can, within an authorized testing framework, transform a vulnerability into an exploitation method, a practice researchers use to materialize the impact of a flaw before it is fixed. Companies like Cisco, Palo Alto Networks, and CodeRabbit are evaluating or already using it for vulnerability research and audits. Despite these uses, identifying a flaw does not guarantee its remote exploitation, which depends on attack conditions, specific configurations, and existing protections. OpenAI presents these offensive capabilities with a desire for regulation and collaborates notably with Dragos. In the lab, models have uncovered two new vulnerabilities related to electrical protection equipment. This gradual approach echoes the caution already observed at Google with Gemini 4 Carbon, initially reserved for trusted researchers and professionals.