OpenAI unveiled GPT-6 Cyber, its latest cybersecurity-focused model, on Thursday. The company says the model is its best so far at finding and fixing software vulnerabilities and following rules set by researchers.
GPT-6 Cyber is the next iteration of GPT-5.6 Cyber and outperforms GPT-6 Astra "on most cyber tasks," according to an OpenAI release.
On ExploitGym, a testing environment for models to target certain vulnerabilities, GPT-6 Cyber scored 48.5% compared to GPT-5.6 Cyber's 32.1% and GPT-6 Astra's 42.4%, according to OpenAI's results.
OpenAI also ran tests on the models' ability to find vulnerabilities in web applications without access to their source code. In this test, GPT-6 Cyber outperformed OpenAI's frontier GPT-6 Astra model by roughly 26%.
Ahead of the release, OpenAI said GPT-6 Cyber was used to find 81 Firefox vulnerabilities that Mozilla classed as "high severity." Mozilla has since patched the flaws.
OpenAI added that GPT-6 Cyber was trained to follow restrictions put on it by researchers. This has been a particular pain point for the company after one of its models broke free of its testing environment and attacked AI cloud provider Hugging Face earlier this year.
GPT-6 Cyber is currently only available through API and Codex to select customers in Daybreak Red, OpenAI's advanced security testing program.