Trust Center
A gateway in front of every call. Here's exactly what we do with it.
Kumo is a billing-and-routing layer — not a data lake. This page lays out, plainly, how your data is encrypted, processed and kept safe: prompt and completion bodies are not retained, while account, billing and usage metadata is kept only to operate the service.
Posture at a glance
What's real, what's next.
| Control | Status | Detail |
|---|---|---|
| Encryption in transit | Live | TLS 1.2+ on every request. |
| Encryption at rest | Live | AES-256 for stored metadata and secrets. |
| Prompt content logging | None | Only billing/usage metadata is retained. |
| Training on your data | Never | Prompts and completions are never used to train. |
| API key storage | Live | Hashed at rest; shown once at creation; revocable. |
| Penetration testing | Roadmap | Scheduled as we exit early access. |
Data flow
Where your prompt goes.
Four steps, end to end — and the prompt is kept only as long as it takes to serve the request.
Your app → Kumo
Your OpenAI SDK calls api.kumo.cloud/v1 over TLS 1.2+. The request carries your prompt to the gateway.
Kumo routes it
Kumo applies routing and metering. We keep usage metadata (tokens, model, timestamp) for billing — not the prompt body.
Forwarded to the provider
The prompt is forwarded to the serving provider (OpenAI, Anthropic or Google) and the provider's response is returned to you verbatim.
Dropped after serving
Prompt and completion are processed to serve your request and nothing more — not stored, not logged, not used for training.
Encryption & keys
Your data stays yours.
Encrypted in transit & at rest
Every request runs over TLS 1.2+; stored metadata and secrets are encrypted at rest with AES-256. API keys are stored hashed and shown once, at creation.
Never trained on
Prompts and completions are never used to train models. They're forwarded to serve your request, then dropped.
Prompt bodies aren't logged
We keep usage metadata for billing — token counts, model, timestamp — not the content of your prompts or responses.
Scoped, revocable keys
Every key is scoped and revocable, shown once at creation.
Sub-processors
Who sees request content.
To serve a request, your prompt is forwarded to the model's provider. These are the only sub-processors that touch request content.
| Provider | Role |
|---|---|
| OpenAI | Model inference (GPT-class) |
| Anthropic | Model inference (Claude-class) |
| Model inference (Gemini-class) |
Questions about how we handle your data?
We answer security questions directly — usually within one business day.