OpenAI has introduced GPT-6 Cyber, a specialized cybersecurity model that can identify software vulnerabilities and develop working exploits, while showing improved ability to stay within the boundaries of authorized security testing.The new model scored 48.5% on ExploitGym, a benchmark that tests whether models can exploit specified vulnerabilities in remote targets. This represents a 51% relative improvement over GPT-5.6 Cyber, which scored 32.1%.OpenAI said GPT-6 Cyber also outperforms its general-purpose GPT-6 Astra model on most cybersecurity tasks, even when Astra does not refuse the requests.The company has classified GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for its predecessor. The classification reflects the model’s stronger ability to understand software, identify vulnerabilities, and develop working exploits.Firefox flaws emerge despite testingOne of the most notable demonstrations of the model’s capabilities came from a vulnerability research campaign involving Mozilla’s Firefox browser.OpenAI researchers used GPT-6 Cyber to uncover 81 previously unknown vulnerabilities that Mozilla classified as high severity. The model also identified 30 previously reported issues that were more serious than initially assessed.The researchers provided Mozilla with evidence, reproducible tests, and proposed patches to help its engineers address the flaws. Mozilla has since begun releasing fixes for many of the reported vulnerabilities.OpenAI also developed a new evaluation for offensive web penetration testing, in which models probe web applications without access to their source code. GPT-6 Cyber achieved 40.3% recall of vulnerabilities, compared with 32% for GPT-6 Astra.Unlike conventional vulnerability assessments that examine source code, the web penetration test required GPT-6 Cyber to identify weaknesses by interacting directly with live web applications. This tests whether the model can discover exploitable entry points using only the behavior and responses exposed by a running service.The model has also demonstrated its ability to reproduce and exploit vulnerabilities in the Linux kernel. It reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. OpenAI disclosed 304 critical findings to Linux maintainers with candidate patches. The company has previously detailed this research separately.Model improves testing boundariesBeyond exploit development, OpenAI has focused on ensuring GPT-6 Cyber respects the limits of authorized security assessments.In its ExploitGym Honeypot evaluation, researchers placed models in isolated environments containing vulnerable supporting services. Although exploiting these services could reveal a task’s flag, doing so was outside the authorized scope.GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment safety prompts. With the prompts enabled, it achieved 100% compliance. GPT-5.6 Cyber scored 94.49% without the prompts.OpenAI said specialized training helps the model follow researchers’ restrictions, remain within authorized testing environments, and report findings accurately.The company is pairing these capabilities with additional safeguards, including cyber-specific monitoring, restrictions on high-risk activity involving critical infrastructure and human oversight for sensitive workflows.GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers. OpenAI plans to expand access to other Daybreak Red customers, while also making GPT-5.6 Cyber available to them.