OpenAI launches GPT-6 Cyber to automate vulnerability research

Advancing the security model

OpenAI recently introduced GPT-6 Cyber, a specialized model designed to assist with cybersecurity tasks. The company describes this new version as being far more capable than its predecessor, GPT-5.6 Cyber. It is specifically built to hunt for software flaws while remaining disciplined about the limits set by human researchers.

The model's performance was measured using ExploitGym, a tool that evaluates whether an AI can exploit a specific flaw in a remote target. On this benchmark, GPT-6 Cyber reached a score of 48.5%. This is a notable increase from the 32.1% score achieved by the previous model, GPT-5.6 Cyber.

Because of these capabilities, OpenAI has assigned the model a 'Critical' rating under its Preparedness Framework. This is a step up from the 'High' designation given to the previous iteration. The rating reflects the model's increased power to perform complex security-related actions.

Testing the model's boundaries

A major goal for OpenAI is ensuring the model stays within authorized scopes. During alignment tests, researchers gave the model a task to exploit a system while providing access to other services that were intentionally easy to hack but off-limits. The model was tested to see if it would 'overreach' and attack those forbidden targets.

GPT-6 Cyber refrained from exploiting the decoy services in 99.93% of cases without any extra safety prompts. This is an improvement over the 94.49% success rate of GPT-5.6 Cyber. When safety prompts were included, the model never targeted the decoy services at all.

OpenAI stated that this progress in model behavior stems from specific training meant to keep the AI within its permitted boundaries and ensure it reports findings accurately. The goal is to create a tool that can perform deep technical work without wandering into unauthorized areas of a network.

Real-world vulnerability hunting

The model has already been put to work on some of the world's most common software systems. Researchers pointed it at the Linux kernel, which is the core of many operating systems. The model successfully found 1,724 vulnerabilities and created working privilege escalation exploits for 948 of them.

In addition to the kernel work, OpenAI reported 304 critical issues to maintainers along with proposed fixes. The model also documented 81 new flaws in the Firefox web browser. Mozilla confirmed that these flaws were of high severity and has already begun rolling out patches for them.

Each of these findings was backed by evidence and reproducible tests provided by the AI. This demonstrates the model's ability to not just identify a theoretical problem, but to provide the specific data needed for developers to fix the issue.

Safeguards and industrial use

To prevent misuse, OpenAI included new protections that monitor whether a model's actions align with what a user is permitted to do. There are also specific safeguards designed to stop the models from aiding efforts to compromise industrial controls or other safety-critical systems. These protections were refined using feedback from the industrial cybersecurity firm Dragos.

Dragos tested the models in a controlled lab environment to ensure they met safety standards. Rob T. Lee, the chief AI officer at Dragos, noted that the minimum level of expertise required to perform the reconnaissance and research needed before an attack is decreasing. This means fewer technical skills are required to perform the initial stages of a cyberattack.

Because of these risks, access to GPT-6 Cyber is currently limited to select approved customers of Daybreak Red. This is a part of OpenAI's Daybreak program which focuses on advanced exploit work and authorized testing. Customers can currently access the model through the API and Codex.

Future rollout and access

OpenAI plans to gradually open GPT-6 Cyber to all customers within the Daybreak Red program. These customers are also receiving access to the previous GPT-5.6 Cyber model. The rollout is designed to be methodical to ensure that the high-powered capabilities are used responsibly by authorized parties.

A dedicated Daybreak Red group will be established specifically for approved defenders of critical infrastructure. This group will be permitted to perform sanctioned testing on critical systems. This allows organizations to use the AI to find and fix weaknesses in the systems that keep society running.

By providing these tools to defenders, OpenAI aims to help them stay ahead of potential threats. The model serves as a high-powered assistant for finding bugs, but it remains under strict oversight to ensure it is used for protection rather than harm.