Security & Trust
What happens to your data, plainly
An authorization layer reads your agents' outbound messages before they ship. That deserves clear answers, not a badge wall. Everything below is split into what is true today and what is planned; nothing here is aspirational dressed as current.
True today
| Area | How it works |
|---|---|
| Where processing happens | Entirely on Cloudflare's edge network. Semantic checks run on Cloudflare Workers AI (Llama 3.3 70B) inside the same infrastructure; submitted content is not sent to OpenAI, Anthropic, or any other model provider. |
| Model training | We never train models on customer content. Per Cloudflare's Workers AI terms, inputs and outputs are not used to train their models either. |
| Encryption | TLS in transit for every request. At rest, records live in Cloudflare D1, R2, and KV, which encrypt stored data. |
| Tenant isolation | Every record is scoped to your agent identity at the query layer: policies, decisions, approvals, and evidence are retrievable only with your key. API keys are stored as SHA-256 hashes and shown exactly once. |
| Audit integrity | Every authorization decision is hash-chained to your previous decision and signed with ECDSA P-256. The public verification key is published at /.well-known/evallayer-signing-key.json, so records can be verified without trusting us. |
| Notifications | Operational emails carry decision metadata only (agent, action type, decision, rule names), never the submitted content. |
| Retention, by data type | Retention is layered so the sensitive part is short-lived and the auditable part survives. Raw action content (the message your agent intended to send) is deleted automatically 90 days after the decision, enforced by a storage lifecycle rule rather than a manual process. Decision records (decision, violation rule names, content hash, policy version, signature) are retained for the life of the account, because deleting them would break the audit chain they exist to provide. Policy versions are retained while the policy is in use. Purging raw content does not weaken verification: the hash chain and signatures still validate. |
| Shorter windows & deletion | Design partners and enterprise accounts can request a shorter raw-content window or zero-retention processing (decision recorded, content never written to storage), configured at onboarding. Deletion of your data on request at any time: email support@evallayer.ai and we complete it within 30 days. |
| Subprocessors | Cloudflare (compute, storage, AI inference), Resend (our operational email), Stripe (payments; card data never touches our systems). Lead-form submissions additionally go to our CRM. |
| Availability | Runs on Cloudflare's global network; live status at /health. No formal SLA yet (see planned). |
| Vulnerability reports | support@evallayer.ai reaches a human. We respond to good-faith security reports and credit researchers who want it. |
Planned, stated honestly
| Item | Status |
|---|---|
| SOC 2 | On the roadmap, sequenced against enterprise demand. Not certified today, and we will not imply otherwise. |
| Formal SLAs | Available as part of enterprise agreements as they are papered; not offered generically yet. |
| Zero-retention mode | Process-and-discard authorization (decision recorded, content never stored) is designed and available to enterprise design partners on request. |
| DPA | Available on request at support@evallayer.ai. |
Questions procurement wants answered that this page missed: email support@evallayer.ai and the answer gets added here.