OpenAI announced the release of GPT-6 Cyber on November 5, a specialized version of its GPT-6 Astra model designed for authorized vulnerability research. This model is available exclusively to a limited number of approved customers in the Daybreak Red tier, which focuses on offensive security tasks such as penetration testing and red teaming.
According to OpenAI, GPT-6 Cyber represents a significant advancement in security research, providing enhanced exploit-finding capabilities while adhering to established rules of engagement. Early findings from the model have already resulted in critical security patches. The model identified 81 previously unknown vulnerabilities in the Firefox web browser, which Mozilla classified as high severity. Mozilla has since implemented fixes for many of these issues.
What the model found
In addition to the new vulnerabilities, GPT-6 Cyber also pinpointed 30 previously reported issues that were more severe than initially assessed. OpenAI researchers provided Mozilla with reproducible tests, evidence, and suggested patches to address these vulnerabilities.
In research concerning the Linux kernel, GPT-6 Cyber reproduced 1,724 vulnerabilities and created working exploits for 948 of them. OpenAI has disclosed 304 critical findings to Linux maintainers, accompanied by proposed patches.
In performance assessments, GPT-6 Cyber achieved a score of 48.5% on ExploitGym, a benchmark for testing models’ ability to exploit specific vulnerabilities in remote targets. This is an improvement from the 32.1% score of the previous model, GPT-5.6-Cyber. In a new web penetration test that involved probing live applications without access to source code, GPT-6 Cyber identified 40.3% of vulnerabilities, up from 32% for GPT-6 Astra.
Built to respect the rules of engagement
OpenAI developed a test called ExploitGym Honeypot, designed to prevent the model from overstepping boundaries during testing, particularly following an incident this summer when its agents unintentionally accessed an external system. The model had a boundary violation rate of 0.07% in graded runs without deployment safety instructions, a substantial reduction from 5.51% for GPT-5.6-Cyber. With the proper instructions, the violation rate dropped to zero.
OpenAI classifies GPT-6 Cyber as Critical, the highest tier in its cybersecurity preparedness framework. Evaluations by Irregular, an AI security lab, suggest that while GPT-6 Cyber is comparable to GPT-6 Astra, it performs better in complex, multi-step tasks.
Who can use it
Access to GPT-6 Cyber is granted through OpenAI’s API and Codex exclusively for approved Daybreak Red customers, who must submit requests from designated networks. OpenAI plans to roll out access to all Daybreak Red customers over the coming weeks, while also making GPT-5.6-Cyber available to all customers within the tier. Additionally, critical infrastructure defenders will receive their own specialized Daybreak Red group.
OpenAI is also expanding early access to its Astra model through Daybreak Blue, which will benefit more organizations and individual security researchers. The company emphasizes the importance of ensuring that gated exploit development leads to faster patch creation compared to unrestricted use of such capabilities.
