OpenAI has made GPT-5.6 Cyber, its security-tuned model, available inside AWS GovCloud (US-East) through its Daybreak Red programme. For American government security teams this matters less as a model launch than as a procurement one: the model now runs in the isolated cloud region many of them are already cleared to use, reached through Amazon Bedrock rather than a separate vendor account.
Daybreak Red is OpenAI's tier for authorised offensive security work: finding vulnerabilities, reproducing them in a controlled setting, validating exploits and drafting mitigations. OpenAI describes the trade as more latitude in exchange for more scrutiny. Verified users get fewer refusals on these tasks, and in return face stronger identity checks, monitoring and access controls. OpenAI also says qualified practitioners should review any finding or proposed fix before acting on it.
Access has two gates. A team needs OpenAI's Daybreak Red approval and separate provisioning from AWS, and OpenAI states plainly that the first does not switch on the second. On data, OpenAI says inference data from Daybreak Red on Bedrock is not used for training, and customers do not have to share their Daybreak Red data with OpenAI. That matches the general Bedrock terms we cover in our Amazon Bedrock review, where inputs and outputs are not passed to model providers.
The model on offer is the older one. In our GPT-5.6-Cyber review we note that OpenAI's launch figures had it completing 95% of advanced exploit-chain and privilege-escalation prompts, against 1.5% for the standard GPT-5.6 Sol. That figure measures willingness to help, not skill. On OpenAI's ExploitGym test, which asks a model to turn a known flaw into a working exploit, GPT-5.6-Cyber scores 32.1% against 48.5% for GPT-6 Cyber, which OpenAI released a day after this announcement. GPT-6 Cyber is not part of the GovCloud release.
Several things remain unclear. OpenAI gives no pricing for the GovCloud offering, no timeline for adding GPT-6 Cyber, and no detail on how eligibility is judged beyond an application form and account team contact. Every capability figure above comes from OpenAI itself and has not been reproduced independently. For agencies, the practical gain is that a capable exploit model can now sit inside an environment their compliance teams already understand, which is often the slowest step in adopting any new tool.