Back to news
launchOpenAI2026-11-06

OpenAI released the cybersecurity-specialized GPT-6 Cyber model on November 5 with a 51 percent relative gain on ExploitGym

GPT-6 Cyber scored 48.5 percent on ExploitGym, up about 51 percent relative to GPT-5.6 Cyber at 32.1 percent, reproduced 1724 Linux kernel vulnerabilities with 948 working privilege escalation exploits and found 81 new Firefox vulnerabilities.

On November 5, OpenAI published the cybersecurity-specialized model GPT-6 Cyber, saying it shows significant improvement over the previous GPT-5.6 Cyber in discovering, validating and fixing software vulnerabilities. The ExploitGym benchmark mainly tests whether a model can exploit a specified remote vulnerability, and GPT-6 Cyber scored 48.5 percent, a relative improvement of about 51 percent over GPT-5.6 Cyber at 32.1 percent. OpenAI also said that without deployment safety prompts, 99.93 percent of tests avoided attacking services outside the authorized scope, and with prompts enabled that figure reached 100 percent.

In practice, OpenAI researchers used GPT-6 Cyber to reproduce 1724 Linux kernel vulnerabilities and developed working privilege escalation exploits for 948 of them. The model also found 81 previously unknown high-severity vulnerabilities in Firefox, and submitted evidence, reproducible tests and candidate patches to Mozilla. OpenAI rated GPT-6 Cyber as Critical for cybersecurity risk under its preparedness framework, up from High for the previous generation.

The model is initially available only to approved Daybreak Red customers, with safeguards including abuse monitoring, credential protection, sandboxes, human supervision and high-risk behavior restrictions for critical infrastructure. Access will gradually expand to all Daybreak Red customers, while GPT-5.6 Cyber is opened to all Daybreak Red customers.

GPT-6 Cyber网络安全ExploitGymDaybreak RedFirefox