Consus puts compliance in the model name for its defense AI gateway
The founder's launch pitch combines government-cloud models behind one endpoint, mandatory compliance suffixes and organization-level controls.
One endpoint, with the data boundary in the request
Consus is pitching aerospace and defense teams an AI gateway that makes the requested compliance level part of the model name. In an October 13 Launch HN post, founder Eric said bare model names are rejected: callers must request a name with a suffix such as :fedramp-high or :itar. The gateway then selects a supported government-cloud backend. The company’s documentation lists AWS GovCloud, Azure Government and GCP Assured Workloads, with availability varying by model and compliance level.
The founder says Consus verifies organizations and puts contracts in place before issuing API keys. Its portal offers model allowlists and user-level model and spending restrictions. The launch post quotes a compliance fee starting at $40 per seat per month, plus model tokens at cost. This is the company’s product and pricing description, rather than a finding that every customer’s use is compliant.
The retention exception in the launch discussion
Consus markets a zero-persistence architecture for prompts and completions, retaining operational metadata for billing and audits. In a reply about provider abuse monitoring, however, the founder disclosed a remaining exception for OpenAI GPT-5.x models on AWS: flagged requests can be retained for up to 30 days. He said the review is automated, stays within AWS GovCloud and is not shared with OpenAI, and that the company is working to disable it. That qualification matters alongside the broader zero-retention wording on the company’s pages.
Equivalency and tool controls have boundaries
Consus says a third-party assessor reviewed it against all 410 FedRAMP High controls. Its compliance page calls this FedRAMP Equivalency; it separately says a Consus-specific DoD IL5 application authorization remains on the roadmap. Those company disclosures do not establish that Consus itself holds every authorization of the underlying clouds.
Its security documentation also distinguishes blocking hosted tools from governing client-executed functions. Certain hosted search and execution tools are rejected or removed, while detected destinations in returned function arguments produce an advisory signal. The gateway does not execute those functions or automatically block every flagged destination; the calling application still has work to do.