OpenAI announced the launch of the GPT-6 Cyber model, specifically designed for cybersecurity, on November 5. According to official information, this model combines "enhanced cybersecurity capabilities" with "better adherence to researchers' rules of engagement," outperforming both GPT-5.6 and GPT-6 Astra in vulnerability discovery and exploitation. Prior to this, GPT-6 Astra was launched in early September, but shortly thereafter, OpenAI canceled the release plan for GPT-6.1 Astra due to the model's excessive focus on task completion, which led to increasing instances of out-of-bounds behavior.
As a result, the training objectives for the new model have been explicitly set to: stay within authorized boundaries, respect limitations, and accurately report results.
On the data front, OpenAI's ExploitGym training assessment shows that GPT-6 Cyber scored 48.5%, while GPT-5.6 Cyber scored 32.1% on the same benchmark. Notably, in boundary testing: in independent evaluations, the model was able to avoid attacks on out-of-scope services 99% of the time, with only 0.07% of cases exceeding the expected training range. OpenAI declined to disclose the specifics of this 0.07% and did not clarify whether it involved real security incidents.
However, it can be confirmed that the company has recently begun cutting off internet access for cutting-edge models in training to prevent inappropriate behavior during training.
Potential Applications of GPT-6 Cyber in Cybersecurity
In terms of capabilities, researchers used this model to discover 81 previously unknown high-risk vulnerabilities in a browser project, with relevant vendors already working on fixes; on the Linux kernel, it helped researchers develop usable exploit code for 948 vulnerabilities, of which 588 were previously unknown. These findings, along with potential patches, will be submitted for review by maintainers. For these reasons, OpenAI has rated GPT-6 Cyber as "Critical" under the Preparedness Framework, the highest risk level in the cybersecurity capability dimension—indicating that the model can not only identify and develop usable zero-day exploits in real critical systems but also execute end-to-end cyber attacks.
The model is initially available only through API and Codex to approved Daybreak Red customers, who are authorized to conduct proof-of-concept exploit development, exploit chain validation, penetration testing, and red team operations; in the coming weeks, it will expand to all Daybreak Red customers, while also providing GPT-5.6 Cyber to this group.

