OpenAI raises GPT-6 Cyber risk rating to Critical after vulnerability tests

A leap in security research

OpenAI has introduced GPT-6 Cyber, a model specifically engineered to assist in security research. The company claims this model is significantly more capable at identifying and fixing software vulnerabilities than previous versions. It is designed to stay within the boundaries set by researchers while performing complex tasks.

The new model showed major gains on key cybersecurity benchmarks. Specifically, it achieved a 51% relative improvement on ExploitGym compared to GPT-5.6 Cyber. It also demonstrated a "step function improvement" when tested on its ability to remain within the scope of a specific task.

OpenAI is using its own internal evaluations to back these claims. The company notes that the model is already being used to secure real-world software. It has uncovered and validated thousands of exploits in critical codebases like Linux and Firefox.

Real-world software impact

The practical application of GPT-6 Cyber has already yielded significant results for major open-source projects. Researchers used the model to reproduce 1,724 Linux kernel vulnerabilities. Of those, the model developed working exploits for 948 of them, which is more than five times the amount produced by the previous GPT-5.6 Sol model.

In the case of the Firefox web browser, the model produced evidence for 81 new high-severity vulnerabilities. It also identified 30 known issues that were actually worse than what had been previously reported. Mozilla began landing fixes shortly after receiving these reports from the researchers.

Beyond the kernel and browsers, the model was used to find a major bug in Brotli, a tool for compressing data that Node.js relies on. This flaw could lead to remote code execution in downstream applications. These findings highlight the model's ability to navigate complex, interconnected software dependencies.

Risk ratings and safety

Because of the model's high performance in creating exploits, OpenAI has raised its cybersecurity risk rating from High to Critical. The company is aware that such power creates risks, such as an agent wandering outside of its assigned task to attack unrelated services. To counter this, OpenAI trained GPT-6 Cyber to respect limits and report results accurately.

In testing, the model avoided unauthorized actions in 99.93% of cases without a deployment prompt. When given extra guidance on safety, it remained within the permitted boundaries during every single test. This dedicated training aims to ensure that the model remains a tool for defenders rather than a weapon for attackers.

OpenAI also monitors suspicious activity to prevent malicious misuse. They check whether user actions match permissions and watch for emerging threats. For example, they monitored attempts to develop exploits against live PHP servers while simultaneously helping defenders fix the underlying flaws.

Protecting critical infrastructure

OpenAI is placing a special focus on protecting critical infrastructure like power plants and water treatment facilities. They launched a $1 billion grant program called Daybreak for Frontline Defenders to help resource-constrained teams. A new group, Daybreak Red for Critical Infrastructure, allows approved defenders to run authorized tests on sensitive systems.

The company also tightened safeguards to prevent the model from providing help that could damage equipment or endanger lives. They are monitoring sensitive-request label rates to ensure the model remains useful for legitimate research without facilitating dangerous actions. They aim to balance high block rates with the need for effective security patching.

Testing on a mock power plant showed the model could separate operational technology from IT noise on its own. It discovered two zero-days in engineering software that handles protective relay data, exposing protection settings and breaker trip logic. OpenAI plans to use this feedback to refine safeguards and form more partnerships with energy companies and national labs.