Security and compliance
How OneType protects your data and your visitors, and what we can and can't claim today.
Your data stays yours
- No training on your data. We don't use your content or your visitors' conversations to train AI models.
- Isolation between customers. Every record belongs to one workspace. Every database and search query is filtered by workspace in code, and results are checked again against the caller's workspace before they are returned. Automated tests write data as one customer and try to read it as another in our test suite.
- Retention.
- Conversations are kept until you delete them. A retention period (30 days to 2 years) exists in OneType, but you can't set it yourself yet; a dashboard control is coming. After that period they are erased; a record without any personal data remains so your statistics stay correct.
- Individual visitors can be exported or erased on request.
- A deleted workspace can be restored for 30 days, then everything is permanently removed.
- Where it runs. In the EU: Hetzner servers in Germany; files and backups on Cloudflare R2 storage in the EU. Our sub-processors are listed at /legal/subprocessors, and we give 30 days' notice before adding one.
Protections built into the product
| Risk | What we do |
|---|---|
| Someone else's site embedding your agent | You list the domains the chat may appear on. The browser itself refuses the chat anywhere else (Content-Security-Policy frame-ancestors). |
| Visitors impersonating your users | Identity comes only from tokens signed by your server with your agent's secret: short-lived, audience-bound, and with rotatable secrets. See Identity. |
| The agent being tricked into actions (prompt injection) | Text from web pages, documents and tools is treated as data, never as instructions. Answers can only link to the sites of the sources they cite, your allowed domains and your own site, so a planted link never reaches a visitor. Any action that changes data needs the visitor's own click in the chat, which the model can't fake. Destructive actions always ask. |
| Data leaking through the chat | Answers are rendered without raw HTML, links are shown with their domain, external images are never auto-loaded, and the chat runs in an isolated frame with a strict Content Security Policy. |
| Crawling or tools reaching internal systems (SSRF) | OneType only connects to public internet addresses, checks every resolved address, and doesn't follow redirects for tools and webhooks. |
| Abuse and cost attacks | Rate limits per visitor, network address, agent and API key. Abusive traffic is never charged to you, and you can set a monthly spend cap. |
| Leaked API keys | Keys are stored hashed and shown once. Restricted keys carry only the permissions you tick, and revoking a key takes effect on the next request. |
| Tampered webhooks and tool calls | Every request we send is signed (Standard Webhooks), with timestamps to stop replays. |
EU AI Act
- Transparency (Art. 50): when a person takes over, the chat says so, and again when the AI answers. Asked directly, the assistant says it is an AI and never claims to be a person. Messages carry
is_ai_generatedin the API and webhooks. - Prohibited and high-risk uses are excluded by our Terms of Service (section 7), unless agreed separately.
Compliance status (honest)
| Item | Status |
|---|---|
| Data processing agreement | Published at /legal/dpa, part of the Terms |
| Sub-processor list with change notices | Published at /legal/subprocessors |
| Encryption in transit (TLS 1.2+) | Yes |
| Secrets (AI keys, tokens) encrypted (AES-256-GCM) | Yes |
| HIPAA | Not offered |
Reporting a vulnerability
Report security issues to security@onetype.ai, with details and steps to reproduce. We read every report. Please don't test against other customers' data.