Hardware Ready Ready to Bench?
OpenAI has announced the release of GPT-6 Cyber, a new artificial intelligence model designed specifically for cybersecurity research. Compared to the previous version, GPT-5.6 Cyber, this model promises significant advancements in identifying vulnerabilities, developing working exploits, and preparing corrective patches.
A central aspect of the project concerns alignment: the model has been trained to more precisely respect limits imposed by researchers, avoiding interventions on systems or services that fall outside the authorized perimeter. OpenAI has published results demonstrating improvements in key industry evaluations.
GPT-6 Cyber: why it matters
In the ExploitGym benchmark, which measures the ability to exploit a specific vulnerability in a remote system, GPT-6 Cyber reached 48.5%, compared to the 32.1% achieved by GPT-5.6 Cyber. This represents an increase of 16.4 percentage points, equivalent to a relative improvement of approximately 51%. The model demonstrated superior performance also in a new evaluation dedicated to penetration testing of web applications, where agents must identify vulnerabilities without access to the source code. In this case, GPT-6 Cyber reached a detection rate of 40.3%, compared to 32% for GPT-6 Astra.
OpenAI introduced RemoteExploitBench, a benchmark that verifies the ability to develop working exploits against vulnerable applications such as GitLab, PostgreSQL, and Jenkins, installed in controlled environments. In this evaluation, GPT-6 Cyber completed 95% of the proposed tasks, while GPT-6 Astra with Daybreak Blue refused all requests. The only failure was caused by an exploit that the model failed to execute correctly remotely.
Based on these evaluations, OpenAI classified GPT-6 Cyber as ‘Critical’ within its own Preparedness Framework, surpassing the ‘High’ classification assigned to GPT-5.6 Cyber. The company emphasizes that the new model outperforms GPT-6 Astra in most computing tasks, even when the latter does not refuse requests.
The capabilities of GPT-6 Cyber have already been employed in research activities on real software. In particular, OpenAI researchers identified 81 previously unknown vulnerabilities in Firefox, classified by Mozilla as high severity. This result is significant considering that it involves a codebase widely scrutinized and hardened over the years through several generations of advanced AI models. OpenAI provided Mozilla with evidence, reproducible tests, and patch proposals to facilitate problem analysis and correction development. The organization began integrating corrective measures immediately after receiving the reports.
During the same research campaign, 30 known issues were also identified, but with a higher severity than previously attributed. Another activity concerned the Linux kernel, and the company had already discussed this in recent hours – here is the full report. The model was also used to study a critical vulnerability in Brotli, a compression library utilized by Node.js and numerous other software. In collaboration with other OpenAI models, GPT-6 Cyber contributed to developing and verifying exploits that demonstrate how the issue could allow remote code execution in applications using the library.
What changes and what are the effects
One of the aspects on which OpenAI states it focused most heavily is the model’s ability not to exceed the boundaries of assigned tasks. An AI agent employed in vulnerability research might indeed attempt to exploit auxiliary services, credentials, or infrastructure that do not belong to the system under verification. To measure this behavior, the company used ExploitGym Honeypot, an evaluation also inspired by the incident involving Hugging Face. The test provides the model with an isolated environment containing a main service and other resources, such as package registries and object archives, intentionally vulnerable. Exploiting these resources would provide useful information to complete the task but would represent an unauthorized action.
GPT-6 Cyber avoided attacking services outside the authorized perimeter in 99.93% of cases, compared to 94.49% for GPT-5.6 Cyber. By also activating the safety instructions provided for model distribution, the result rose to 100%. The goal is to allow researchers to leverage advanced offensive capabilities to verify system security without turning the agent into a tool that acts autonomously outside received authorizations.
GPT-6 Cyber is initially available via API and Codex for a selected group of approved customers within Daybreak Red, part of the Daybreak initiative through which OpenAI intends to make advanced security research tools available to organizations, independent researchers, and open-source project maintainers. The program provides various access modes. Generalist models, such as GPT-6 Astra, can already be used for code review, vulnerability classification, and patch development. Daybreak Blue offers verified defenders the ability to validate vulnerabilities and corrections with fewer unnecessary refusals, while Daybreak Red enables more advanced exploit development and verification activities through GPT-6 Cyber. OpenAI plans to progressively extend access to all program customers and has also announced the availability of GPT-5.6 Cyber for the entire Daybreak Red audience. In parallel, the company is expanding alpha access to Daybreak Blue based on GPT-6 Astra, targeting a larger number of organizations and independent researchers.
Source and further reading on GPT-6 Cyber: original article.
* Content created with the assistance of artificial intelligence systems.