OpenAI Unveils Security-Specialized "GPT-6 Cyber" ... 48.5% Success Rate in Vulnerability Verification Benchmark

OpenAI unveiled "GPT-6 Cyber," an artificial intelligence model specialized in cybersecurity, on the 5th (local time). The model is designed to strengthen capabilities in discovering and verifying software vulnerabilities and developing security patches.

GPT-6 Cyber can identify vulnerabilities in large codebases, verify whether they are exploitable in practice, and generate patch candidates and analysis materials for fixes. OpenAI said the model is focused on helping security teams proactively find and remediate vulnerabilities.

OpenAI also said it placed emphasis on improving the model's ability to comply with approved task scopes and enhancing safety for security research. To that end, it applied purpose-built training so the model would follow the scope and limitations approved by security researchers.

GPT-6 Cyber is initially being provided through API and Codex to a limited number of approved Daybreak Red customers. OpenAI plans to expand access to all Daybreak Red customers in the future.

According to OpenAI, GPT-6 Cyber posted a 48.5% success rate on its vulnerability verification evaluation.

The company presented case-by-case examples of security research conducted with GPT-6 Cyber. In work on the Mozilla web browser Firefox, it identified 81 previously undisclosed high-risk security vulnerabilities and generated reproducible validation materials for each one, which it delivered to Mozilla. Mozilla then deployed fixes for multiple vulnerabilities.

In Linux kernel research, it reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. OpenAI reported 304 major findings and patch candidates to maintainers in line with Linux community guidelines. Additional vulnerability remediation work is also under way.

In performance comparisons, GPT-6 Cyber's success rate on ExploitGym, a metric for assessing exploitability, was 48.5%, higher than GPT-5.6 Cyber's 32.1%. The results were presented as evidence of GPT-6 Cyber's superior performance. Photo credit: OpenAI.

OpenAI presented results showing GPT-6 Cyber outperformed previous models in performance evaluations. In the "ExploitGym" benchmark, which measures whether known vulnerabilities can be exploited in practice, GPT-6 Cyber posted a 48.5% success rate, while GPT-5.6 Cyber posted a 32.1% success rate. As a result, GPT-6 Cyber held about a 16 percentage point lead over GPT-5.6 Cyber on ExploitGym.

OpenAI also disclosed initial results from "RemoteExploitBench," which measures the ability to carry out complex security tasks on servers that simulate real-world environments. In that evaluation, GPT-6 Cyber completed 95% of the tasks.

Given the risk that cybersecurity AI agents may access unauthorized systems and services during vulnerability research, OpenAI applied specialized training to GPT-6 Cyber and strengthened safeguards to prevent misuse. The specialized training included complying with approved targets and task scopes, following restrictions, and accurately reporting results. The enhanced safeguards consisted of blocking high-risk cyber activities, detecting misuse, and monitoring whether model behavior remained within the scope approved by users. They were designed to limit requests that could harm critical infrastructure or essential services while supporting approved security research and vulnerability verification.

OpenAI said it used GPT-6 Cyber in security testing of its own infrastructure, and explained that this led to vulnerability fixes and improved restrictions on network access.

Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=242738

References

• https://www.itdaily.kr/news/articleView.html?idxno=242738