Security, Privacy & Data Handling
SafePrompt is a security product, so the burden is on us to be specific. This page describes what we actually do today. Where something is not in place, it says so rather than describing a target state.
Last reviewed: 21 August 2026
In short: validating a prompt means sending it to a language model, so your prompt text reaches our inference providers. Safe prompts are not stored. Blocked prompt text and the client IP are held for 24 hours and then anonymised by an hourly job. We do not sell data. We have not had a third-party penetration test.
What we store
- Safe prompts: prompt text is processed in memory and is not written to storage. Only the result and request metadata are logged.
- Blocked prompts: prompt text and the client IP are retained temporarily for network defence, then anonymised.
- Request metadata: timestamp, hashed API key, verdict, threat categories, processing time.
- Account data: email and a hashed password. Billing is handled by Stripe and we do not store card numbers.
- Attack pattern hashes: retained indefinitely. These are cryptographic hashes and carry no personal data.
Retention and deletion
| Data | Retention | Enforced by |
|---|---|---|
| Blocked prompt text | 24 hours | Hourly retention job |
| Client IP (raw) | 24 hours | Hourly retention job |
| Multi-turn session data | 2 hours | Hourly session cleanup |
| Request metadata | Rolling window | Scheduled deletion |
| IP and pattern hashes | Retained | No personal data |
| Account data | Life of account, then 90 days | Deletion request |
How this is enforced, and how it failed before. The retention job previously reported success while deleting nothing, because it called a database function that did not exist in production. Job status alone was not evidence. The job now asserts the outcome: it queries for personal data older than the retention window after each run and raises an error if any is found, so a silent no-op fails loudly instead of reporting success.
Who else processes your prompts
Validating a prompt means sending it to a language model, so prompt content leaves our infrastructure. The companies that receive it are listed here and on the privacy page.
- OpenRouter: routing gateway.
- Groq, Cloudflare, Together AI, DeepInfra: inference providers.
- Supabase: database, hosted on AWS.
- Vercel: API hosting, AWS us-east-1.
- Cloudflare Pages: frontend delivery.
- Stripe: payments. Stripe does not receive prompt content.
Provider routing is pinned in code and versioned in git, so a prompt cannot reach a provider outside this list, and any change is a reviewable code change rather than a silent configuration switch. Inference providers operate globally and the region for a given request is not pinned. If you need a fixed inference region, contact us before integrating.
Infrastructure and keys
- All API traffic uses TLS 1.2 or higher, with HTTPS enforced.
- The database is encrypted at rest with AES-256.
- Customer API keys are stored hashed, so we cannot retrieve your plaintext key.
- Provider tokens and database credentials are platform-managed secrets and are never committed to source. A pre-commit and pre-push scanner blocks live-format keys.
- Customer data is scoped by API key, with no cross-customer access.
- We do not operate a dedicated hardware security module or external key-management service.
What we have not done yet
These are the honest gaps. They are listed because a security vendor that only publishes its strengths is not giving you enough to evaluate it.
- No third-party penetration test. None has been performed to date. A first external test is targeted once paid usage supports the cost. Next review 1 October 2026.
- No SOC 2. Not started, reviewed quarterly. Next review 1 October 2026.
- No status page or availability SLA. Design your integration with an explicit fallback. The API fails closed on internal errors.
- No externally curated benchmark yet. Our detection benchmark is written in-house, so it is a regression instrument rather than independent evidence. See limitations.
- Access review is informal. Production access is limited to a small founding team. Administrative changes to block and allow lists are logged, but a formal periodic access review is not yet in place.
Your rights
- Access and export: available from your dashboard.
- Deletion: delete your account and associated data from dashboard settings.
- Data Processing Agreement: available on request from [email protected].
- We do not sell personal information.
- Opt out of network intelligence: available on paid tiers.
Reporting a vulnerability
Report security issues to [email protected]. We commit to acknowledge within 48 hours, give an initial assessment within 5 business days, credit you in the fix announcement if you want it, and not pursue legal action against good-faith research.
Privacy and DPA requests: [email protected]. General support: [email protected].