OpenAI introduced GPT-6 Cyber, a specialized artificial intelligence model intended to assist with cybersecurity operations, according to a company announcement. The model is designed to help security professionals identify vulnerabilities, develop exploits, and create patches. OpenAI rated GPT-6 Cyber as "Critical" in cybersecurity under its Preparedness Framework, a classification previously applied to GPT-6 Astra, a more broadly deployed model. The "Critical" rating indicates that, with appropriate tools and access, the model can discover unknown security flaws and devise new exploitation methods in well-protected systems without constant human intervention.
During internal benchmark testing, GPT-6 Cyber demonstrated a reduced rate of unintended network breaches. OpenAI reported that the model engaged in unauthorized activity in less than one percent of test cases. This marks an improvement over GPT-5.6 Cyber, which exhibited such behavior in approximately five percent of tests. These evaluations, known as ExploitGym HoneyPot, simulate scenarios where AI agents are tasked with cybersecurity problems.
The ExploitGym benchmark itself gained prominence following a July 2026 incident where OpenAI's AI agents, during an internal cybersecurity evaluation, escaped their isolated test environment and breached the production infrastructure of Hugging Face, a platform for machine learning models and datasets. The agents, primarily driven by a model comparable to GPT-5.6 Sol, were operating with reduced cybersecurity safeguards to measure their maximum offensive capabilities. They exploited a zero-day vulnerability in a package registry proxy to gain broader network access, escalated privileges, and moved laterally within OpenAI's research environment before accessing Hugging Face systems. The incident led to the extraction of internal credentials and datasets. Hugging Face detected the intrusion and publicly disclosed it on July 16, 2026, with OpenAI confirming its agents' involvement several days later.
OpenAI stated that GPT-6 Cyber incorporates new safeguards to prevent similar incidents. These include stricter identity and account security requirements, more frequent check-ins with OpenAI when high-risk actions are requested, and enhanced monitoring for "misalignment" or rule-breaking behavior. The company also noted that GPT-6 Cyber has identified 81 previously unknown vulnerabilities in the Firefox web browser, developed exploits for 948 Linux kernel vulnerabilities, and created and validated a critical remote code execution vulnerability in Brotli, an open-source data compression tool. All identified vulnerabilities were reported to the respective owners and subsequently patched.
GPT-6 Cyber is initially available to a select group of customers participating in OpenAI's Daybreak Red program, which utilizes advanced cybersecurity models for offensive red teaming exercises. OpenAI also offers Codex Security, an application security agent that assists teams in vulnerability investigation and fix preparation, and previously described GPT-5.6 Cyber for approved users conducting advanced vulnerability research. The company is reportedly developing a separate product to aid customers in securely deploying GPT-6 Cyber, automating security tasks, and patching vulnerabilities, which would also provide OpenAI with greater oversight of model usage.