Private AI for security policy drafting
An early security policy draft reads less like a policy and more like an inventory of what's missing: systems with no assigned owner, the place the access review breaks down, the vendor that never got a security questionnaire. Paste that draft into a shared AI chat service to tighten the wording, and an outline of the company's open security gaps now sits in that service's logs. Running the drafting model on a dedicated machine keeps that inventory out of a shared product.
What a policy draft reveals
A published policy describes the controls a company has decided to run. The draft that preceded it usually describes the process of figuring that out, which means it names the systems without a control yet, the access reviews that haven't happened in a year, and the exceptions nobody's tracked. That's a more useful document to an attacker than the finished policy, and it's the document a security or compliance lead pastes into an AI tool to word a control statement or check it against a framework.
Deadline pressure is the reason AI helps here (a policy has to exist before the audit finding closes) and also the reason nobody stops to ask where the draft ends up.
What changes when the drafting model is self-hosted
A dedicated DGX Spark running the policy-drafting model keeps every draft, gap list and control mapping on one machine assigned to the team, instead of in a chat product shared with millions of other accounts. The team still gets a model that can turn a control requirement into policy language, check a draft against a framework's structure, and flag a section that is missing a control the framework expects.
Open WebUI handles the drafting interface. Load the framework text or the company's existing control inventory as reference material, using the retrieval approach in the on-premise RAG guide, and the model checks a draft against real requirements instead of a general sense of what a policy should say. For the audit itself, see private AI for compliance audit prep; if DevOps runbooks feed into the same policy work, see private AI for DevOps runbooks.
What the model is useful for, and what stays with the security lead
A drafting model turns requirements into readable policy language and to catch a missing control section before a reviewer does. Deciding what the actual policy should require, which risk to accept versus remediate, and whether a control is in place stays with the security lead who owns the program. A well-worded policy that doesn't reflect what the company does is worse than no policy, since it's the version an auditor or a plaintiff's attorney will hold you to.
What this doesn't solve
Self-hosting the drafting model doesn't close the gaps a draft reveals, and it doesn't pass an audit on its own. What it removes is one specific exposure: a map of the company's weak points sitting in a shared AI service's history while those weak points are still open. The machine is still hosted by GPUwerk, which keeps infrastructure admin access and, under the DPA, does not access workload content. See pricing for what a dedicated Spark costs.