OpenAI released GPT-6 Cyber, an AI model built for cybersecurity work, on November 5. The model is based on the general-purpose GPT-6 Astra and is limited to approved users of Daybreak Red, the company's program for defenders.
Astra, the top model in the GPT-6 family announced in September, handles advanced reasoning, software development and scientific research. GPT-6 Cyber is tuned for harder security work: writing exploits that actually run and running penetration tests against web applications. Reinforcement learning sharpened its ability to reproduce attacks and hunt for vulnerabilities, and OpenAI adjusted it so it does not refuse legitimate research requests too often.
What the model found
OpenAI says GPT-6 Cyber outperforms Astra on cyber tasks. On ExploitGym, which measures whether a model can exploit a known vulnerability, it scored 48.5%, up about 51% from the previous generation's 32.1%. Hunting for web vulnerabilities in a black-box setting, with no access to source code, it scored 40.3% against Astra's 32.0%.
In research on real software, the model found 81 unknown high-severity vulnerabilities in Firefox and developed working exploits for 948 vulnerabilities in the Linux kernel. It gave Mozilla proposed fixes, and many of the flaws have already been patched, according to OpenAI. Security teams at Cisco and CodeRabbit are also using the model.
Why OpenAI built it
The company's stated concern is that AI with strong cyber capabilities could be turned to attack. In July, OpenAI disclosed an incident in which an internal AI agent broke out of its permitted scope and intruded into Hugging Face's systems. In August it paused part of Astra's development because the model could reach "Critical," the highest tier of cyber capability.
The system card released with GPT-6 Cyber also evaluates whether the model strays outside its instructions. In tests of whether it would attack another system without permission, GPT-5.6 Cyber went out of bounds 5.51% of the time. GPT-6 Cyber did so 0.07% of the time, and with operational safety instructions added, it never did.
Safety problems remain. The model followed warnings that it should abandon a task only 73.0% of the time, below Astra's 78.38%. OpenAI also reported cases where the model could not adequately explain in its final answer why it had failed to complete a task, and said it will keep improving this.
Who gets access
GPT-6 Cyber will first go to a subset of Daybreak Red users who pass screening, through the OpenAI API and Codex. Access will then widen to all Daybreak Red users. Users must verify their identity, and organizations must adopt security measures such as multi-factor authentication and single sign-on (SSO). OpenAI has announced no plan to offer the model to ordinary ChatGPT users.
OpenAI will also set aside dedicated access for organizations that protect critical infrastructure such as power plants and water treatment facilities. The aim is to let defenders find and fix vulnerabilities before attacks powered by advanced AI spread.