Factual posture for compliance officers and procurement. What is implemented, what is in progress, and what depends on your deployment tier.
Data inspected for redaction stays on the appliance. Sanitised prompts go to your chosen LLM provider — local models on the box, Trinito Cloud, or public APIs under your own keys. Uploaded files are stored locally; only sanitised extracted text is sent to an LLM. Audit logs stay on the appliance and are exportable on demand.
Trinito is engineered against the controls of recognised privacy, security and AI governance standards. We do not yet hold third-party certifications — that is on the roadmap once the product is in the hands of paying customers. We use "aligned" where controls are implemented but not yet independently certified, and "in progress" where external assessment is booked.
For procurement teams: we maintain a Statement of Applicability mapping each of the standards above to specific appliance controls and operating practices. Email security@trinito.com and we will return it within two working days, with certification status clearly marked.
Trinito's central infrastructure never sees customer prompts or responses. The appliance routes every prompt directly to the chosen LLM — local models on the box, Trinito Cloud, or your own provider keys — using credentials in its local encrypted secrets store.
Our licensing server receives a daily check-in with the appliance ID and returns signed config: active subscriptions, API keys for Trinito Cloud, monthly caps. No prompt content, no response content, and no audit log ever leave your office unless you export them.
All secrets on the appliance — provider API keys, session tokens, audit-log signing material — are encrypted at rest with libsodium secret-box, using a per-appliance master key. All inter-component traffic uses TLS 1.3. Signed config blobs from the licensing server are verified against a baked-in Trinito public key before any contained keys are accepted.
Every appliance ships with a unique public hostname under box.trinito.com and a publicly-trusted TLS certificate issued via ACME DNS-01. Browsers on your LAN see a real HTTPS certificate — no self-signed warning workflow and no manual trust configuration.
Certificates renew automatically through the licensing server every sixty days; private keys remain on the appliance.
An append-only, hash-chained audit log records every prompt processed, every redaction decision, and every external send. The log stores cryptographic hashes of prompt and response content — not the content itself — so we can evidence what happened without retaining the underlying personal data.
Each entry includes a cryptographic link to the previous row; altering history breaks the chain and is detectable on export. The database enforces append-only behaviour via a row-level trigger, so even root cannot silently rewrite past entries without breaking the chain. Compliance teams can export from the appliance on demand.
Conversation history — which does retain prompt and response text for user reference — lives in a separate per-user store on the appliance, with per-conversation deletion controls and full erasure on user request.
Our first independent penetration test is booked with a CREST-accredited firm; we will refresh annually thereafter. Once the first report is in hand, an executive summary will be available under NDA to enterprise prospects on request.
Report security issues to security@trinito.com. We aim to acknowledge reports within one working day and will publish a full disclosure policy and PGP key here when ready.
The list below is complete for Trinito-operated services. If you use bring-your-own keys, the appliance talks directly to your chosen provider under your contract — we are not in that data path.
| Processor | Purpose | When it applies |
|---|---|---|
| Cloudflare (UK / global) | DNS for box.trinito.com per-appliance hostnames and ACME DNS-01 challenges for per-appliance TLS certificate issuance |
Always (Trinito Compact and Trinito Standard tiers) |
| Trinito-owned bare-metal infrastructure | Trinito Licensing Server — daily signed configuration blobs, no prompt or response data | Always (Trinito Compact and Trinito Standard tiers) |
| Ollama Cloud | LLM inference for the Trinito Cloud model entry | When the customer chooses Trinito Cloud as their LLM provider |
| OpenAI | ChatGPT models | When the customer adds an OpenAI key under Bring-Your-Own; direct from appliance under their contract |
| Anthropic | Claude models | When the customer adds an Anthropic key under Bring-Your-Own; direct from appliance under their contract |
| Gemini models | When the customer adds a Google key under Bring-Your-Own; direct from appliance under their contract |
Customers using Bring-Your-Own provider keys: Trinito is not a sub-processor on that route. The appliance talks directly to your chosen provider under your contract with them.
We maintain an incident response process for security and availability events affecting customer appliances or Trinito-operated infrastructure. Critical incidents (customer-visible degradation or suspected data exposure): acknowledgement within one hour during UK business hours, or within four hours outside them for Sovereign support customers. High severity (single appliance or service affected, no confirmed exposure): four-hour acknowledgement. Medium and low: next UK working day. We provide weekly written updates until resolution for confirmed security incidents.
We will answer your questionnaire with sourced, dated responses — or walk through the Sovereign air-gap deployment on a call.