Press "Enter" to skip to content

OpenAI unveils GPT-6 Cyber, claims record‑breaking vulnerability detection

OpenAI announced the launch of GPT-6 Cyber, a new artificial‑intelligence model built specifically for cybersecurity research. The company said the model is capable of locating software weaknesses, creating test cases, and assisting security teams in developing patches. In its first public demonstration, GPT-6 Cyber reported 81 previously unknown high‑severity vulnerabilities in Mozilla’s Firefox browser and produced functional exploits for 948 flaws in the Linux kernel.

Firefox findings and collaboration with Mozilla

According to OpenAI, Mozilla classified all 81 newly discovered Firefox vulnerabilities as high severity. OpenAI provided reproducible test evidence, detailed proof‑of‑concept exploits, and suggested patches to Mozilla’s security team. Mozilla began integrating fixes shortly after receiving the reports. In addition to the 81 novel bugs, the model also highlighted 30 known Firefox issues that were more severe than earlier assessments had indicated.

Linux kernel exploits and benchmark performance

In the Linux arena, GPT-6 Cyber reproduced a total of 1,724 known vulnerabilities and successfully generated working exploits for 948 of them. OpenAI noted that this output exceeds the 166 exploits produced by its earlier GPT-5.6 Sol model when tested under the same conditions, representing more than a five‑fold increase.

The company disclosed 304 critical findings to kernel maintainers, accompanied by candidate patches, while work on additional fixes continues. Performance on established testing suites also improved markedly. On the ExploitGym benchmark, GPT-6 Cyber achieved a score of 48.5 percent, up from 32.1 percent for GPT‑5.6 Cyber – a relative improvement of 51 percent.

On a newly introduced web penetration‑testing benchmark, the model recorded a recall of 40.3 percent, compared with 32 percent for the GPT‑6 Astra variant. In the RemoteExploitBench assessment, GPT‑6 Cyber completed 95 percent of the tasks, whereas GPT‑6 Astra with the Daybreak Blue configuration refused all tasks. OpenAI explained that the single task GPT‑6 Cyber failed involved an exploit that could not be executed remotely.

Safety mechanisms and infrastructure testing

OpenAI emphasized enhancements to the model’s safety controls. In tests that did not include a deployment prompt, GPT‑6 Cyber avoided exploiting services outside the authorized scope in 99.93 percent of cases, and it achieved 100 percent compliance when a prompt was provided. By contrast, its predecessor, GPT‑5.6 Cyber, stayed within scope in 94.49 percent of prompt‑less tests.

The firm classified GPT‑6 Cyber as “Critical” under its Preparedness Framework for cybersecurity capabilities, upgrading from the “High” classification given to GPT‑5.6 Cyber.

Independent evaluation by Dragos, a global provider of extended operational technology (xOT) cybersecurity, further illustrated the model’s potency. In a controlled laboratory that simulated a power‑plant environment, Dragos used the cyber‑permissive Daybreak Red configuration of GPT‑6 Cyber and discovered two previously unknown vulnerabilities in engineering software that manages protective relay data. OpenAI responded by refining the model’s safeguards following the test.

OpenAI said it has introduced a suite of protective measures, including cyber‑specific monitoring, threat monitoring, sandboxing, human oversight, and additional protections for critical infrastructure. Access to GPT‑6 Cyber will initially be provided through OpenAI’s API and Codex to a limited group of approved Daybreak Red customers, with a gradual expansion planned. Daybreak Red is described as a configuration for advanced exploit development and authorized testing, while Daybreak Blue is intended for vulnerability validation and fix development.

To support frontline defenders, OpenAI announced a USD 1 billion Daybreak grant programme aimed at organizations with limited resources, such as government bodies and essential service providers. The company also created a separate subgroup for approved critical‑infrastructure defenders to conduct authorized security testing.

The announcement, dated November 6, 2026, marks a significant milestone for AI‑driven security research, but it also raises concerns about potential misuse of a model that can autonomously generate functional exploits at scale.

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *