# security

how the service is secured, and how to report a problem.

updated 2026-11-15

> draft: not reviewed by a lawyer. do not rely on it before launch.

## what we do

- Tenant isolation: the organization comes from the verified key or sign-in, never from a request field. Every public database function is covered by a cross-tenant test matrix.
- Keys are hashed, scoped, expiring if you set it, and rate limited. Failed attempts are rate limited per source.
- Recipes are signed; the signature and the content hash are verified before anything runs. Origins are pinned to the recipe's declared list and checked on every redirect hop.
- Page content is untrusted. Extraction returns only typed fields; nothing from a page is treated as an instruction. Secrets never enter model context and are redacted from logs and traces.
- Writes are denied by default. Approvals are single use, expiring and bound to the payload hash. The identity that requested a run cannot approve it.
- The ledger is hash-chained and verifiable offline.

## known gaps

See the threat model in the repository for the current list. Examples: a hostname that later resolves to a private address is not caught at webhook registration; the failed-attempt limiter keys on a source address that a caller can influence.

## reporting

Send reports to the security contact (to be published before launch). Please give us time to fix an issue before disclosing it.
