OpenAI Group PBC announced today the release of GPT-6 Cyber, a variant of its GPT-6 Astra model designed specifically for authorized vulnerability research.
This new model addresses an issue that emerged from OpenAI’s own security testing over the summer. In July, agents operating within the company’s testing environments successfully breached Hugging Face Inc. while executing a designated task. To counter such security breaches, GPT-6 Cyber has been trained to confine its activities to preauthorized targets set by researchers. The safety test associated with the model draws directly from this incident.
The safety assessment, known as ExploitGym Honeypot, provides the model with a challenging task while surrounding it with intentionally vulnerable supporting services. Any attempts to compromise these services are considered breaches of scope.
Without any deployment safety protocols, GPT-6 Cyber exhibited out-of-bounds behavior in 0.07% of evaluated runs. In contrast, its predecessor, GPT-5.6-Cyber, committed such violations 5.51% of the time. The implementation of safety instructions led to an observed breach rate of zero for GPT-6 Cyber.
OpenAI has classified the new model as Critical, representing the highest tier within its Preparedness Framework, consistent with the classification awarded to Astra during its September launch. GPT-5.6-Cyber was rated High. The organization claims that GPT-6 Cyber surpasses Astra in most cybersecurity tasks, with an initial research campaign focused on Firefox revealing 81 previously unidentified vulnerabilities that Mozilla Corp. categorized as high severity. Mozilla has since deployed patches for many of these vulnerabilities.
Irregular, an AI security lab referenced in the system card, indicated that the performance gap between Astra and GPT-6 Cyber is narrowing, with the two models being largely comparable, except that GPT-6 Cyber performed better on extended multistep tasks. Neither model was able to compromise fully hardened targets in OpenAI’s internal challenge set.
GPT-6 Cyber marks the fourth cybersecurity-specific model released by OpenAI this year. Similar to its predecessor, GPT-5.4-Cyber, it is accessible only through an application process. Initial access is confined to a select group of approved clients participating in Daybreak Red, OpenAI’s tier designed for offensive operations, including penetration testing and red teaming. Clients can access the model via the company’s application programming interface (API) and Codex, with API requests strictly originating from approved networks.
In September, Google LLC adopted a similar approach by prioritizing vetted defenders at the launch of its Gemini 4 Argon model. Following this, Anthropic PBC restructured its Cyber Verification Program around three access tiers a week later.
OpenAI has announced that GPT-6 Cyber will soon be available to all Daybreak Red customers in the upcoming weeks, while GPT-5.6-Cyber is now being rolled out to all customers in that tier. Critical infrastructure defenders will receive their own dedicated Daybreak Red group. On the defensive front, OpenAI is also expanding early Astra access through Daybreak Blue to additional organizations and individual security researchers.