OpenAI announced GPT-6 Cyber on November 5, saying the model helped develop working exploits for 948 Linux kernel vulnerabilities and identify 81 newly discovered high-severity Firefox flaws. The company is limiting initial access to approved Daybreak Red customers rather than releasing it widely.

These are OpenAI’s reported research results, not evidence of 948 attacks in the wild. The findings could help defenders find and fix bugs. They also raise a difficult question: who should be allowed to use a tool that can help build exploits?

What OpenAI says the model found

OpenAI says its researchers reproduced 1,724 Linux vulnerabilities and developed working exploits for 948 of them. It reported that 304 critical Linux findings were sent to maintainers with candidate patches.

  • 1,724 Linux vulnerabilities reproduced
  • 948 working exploits developed
  • 304 critical Linux findings sent to maintainers with candidate patches
  • 81 newly identified Firefox vulnerabilities classified as high severity by Mozilla

OpenAI said Mozilla had started fixing many of the Firefox flaws. Reuters also reported OpenAI’s claim that 588 of the Linux vulnerabilities were previously unknown. The company says people reviewed the Linux findings before disclosure.

A higher benchmark score, with limits

On ExploitGym, a test of whether a model can exploit a specified vulnerability, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber, according to OpenAI. That measures performance on a defined benchmark. It does not show how often the model could break into real-world systems.

OpenAI’s system card classifies GPT-6 Cyber as “Critical” for cybersecurity under its Preparedness Framework. The company also says the model stayed within authorized scope more often than its predecessor during tests. Those results describe test conditions, not a guarantee of behavior in every deployment.

Access is expanding, but not public

At launch, OpenAI said GPT-6 Cyber would be available through its application programming interface and Codex to a select group of approved Daybreak Red customers. It planned to expand access to other approved customers over the following weeks. Safeguards include account protections and monitoring for use beyond a customer’s authorization.

For organizations that rely on Linux, the potential benefit is faster vulnerability research and candidate fixes. The risk is that exploit-building capability could spread beyond authorized use as access grows. OpenAI’s announcement describes its safeguards, but the reported results do not establish how well those controls will hold up at wider scale.

The reported discoveries show both sides of the technology: a potential boost for finding flaws, and a powerful capability OpenAI is choosing not to release broadly.