Security and Reliability
This page answers the questions a security or risk team asks before putting Auditail in front of money movement. If something you need isn't here, email [email protected] with "Security" in the subject.
When Auditail is down or slow
| Situation | What happens |
|---|---|
Server-side policy (serverPolicy: true) and Auditail can't be reached | guard() denies the action (code: "policy_unavailable"). It makes up to 3 attempts, each with a 10-second timeout, before giving up. |
| An approval is needed and Auditail can't be reached | Denied. An action that needs approval never runs without one. |
| A rule in your code throws an error | Blocked immediately. It never falls back to allow, and never pages a human. |
| Local rules only (no server policy) | Rules run in your process. Event delivery happens in the background; a delivery failure never blocks or allows anything. |
| Receipt redemption can't reach Auditail | redeemed: false, code: "unavailable". Don't execute. |
Latency. A server-side policy check is one HTTPS call to api.auditail.io, which runs on Cloudflare's network. In November 2026 we measured a median of 349 ms and a 90th percentile of 394 ms for POST /v1/decide, from a client in India. Most of that is database round trips. Rules that run in your process take under 5 ms. If that latency is too high for your flow, run the check before the user-facing step, or tell us your target; reducing it is on our list.
Throughput and contention. Each account's audit log is a single hash chain, so writes to one account are serialized. A write that loses a race retries automatically, up to 12 times with growing random delays, and only then returns 503 contention. A policy check returns 503 too, so the SDK retries it, and if it still fails the action is denied. Accounts don't affect each other. What we measured on production in November 2026, for simultaneous POST /v1/decide calls on one account:
| Simultaneous decisions | Recorded | 503 | Median | Slowest |
|---|---|---|---|---|
| 10 | 10 | 0 | 1.0 s | 1.9 s |
| 25 | 25 | 0 | 1.6 s | 3.5 s |
That's roughly 7 decisions per second per account under a burst. Nothing is lost, but bursts are slow, and for Stripe card authorizations a decision that takes longer than 1.5 seconds is declined. If your agents act in bursts, spread them over time or use separate accounts per agent group, and tell us your peak rate. Raising this limit is on our list.
Uptime. There is no uptime commitment during the private beta. Status and incidents are communicated by email to design partners.
Where data is stored
- Hosting: Cloudflare Workers. Database: Cloudflare D1, located in Cloudflare's Asia-Pacific (APAC) region. If you need data to stay in the EU or US, tell us, or self-host in your own Cloudflare account.
- Encryption: TLS in transit. Cloudflare encrypts D1 data at rest with AES-256 (Cloudflare documentation).
- What we store: for each action, the agent id, action name, the payload as the SDK sent it, the decision and reason, and any context you send (
intent,userRequest,source). Also approvals, policies, freezes and receipt redemptions. - Payment provider integrations store the provider's webhook signing secret (it verifies requests and can't move money), and from each request only what the policy needs: amount, currency, merchant details, card id and last 4 digits, beneficiary name and country, and a short hash instead of the bank account number. We never ask for, or store, payment API keys.
- What we don't store: secret fields. The SDKs replace values of keys like
password,token,secret,apiKey,authorization,ssnandcardwith[REDACTED]before the event leaves your process, and you can add your own keys (redactKeys). To keep more out of Auditail, send only the fields your policy needs: for example a beneficiary id rather than the full bank details. - Retention: during the beta, events are kept until you ask us to delete them. Cloudflare's point-in-time recovery keeps database history for up to 30 days after a deletion.
Credentials and access
- Two kinds of credential. Agent keys (
atl_live_) can request decisions and approvals, but can never approve, change the policy or lift a freeze. Approver tokens (atl_appr_) are issued per person. - Act-only agent keys. An agent key can be restricted so it can't read the audit log, exports, reports, policy or connections either (details). We recommend it for every agent. Approver tokens are bearer tokens issued per person; there's no SSO or MFA yet.
- Stored as hashes. API keys and approver tokens are stored as SHA-256 hashes. We can't show them again, and we can revoke them.
- Signing key. The Ed25519 key that signs receipts, checkpoints and reports is a Cloudflare Worker secret. It isn't in the code, the repository or the database.
- Dashboard. It keeps the credential in the browser tab's session storage only and sends it only to Auditail. The page is served with a strict Content Security Policy (no third-party scripts).
- Logs. The server logs errors without payloads.
Agents that read untrusted content
- Provenance and session taint. Mark where data came from (
source,fieldSources) and which sessions have read untrusted content (markUntrusted(), or the MCP proxy'suntrustedtools). Money-moving actions in a tainted session need a human by default. Auditail can't see inside the model: this is only as complete as your integration's marking, and it's deliberately coarse. See session taint. - Only trusted reporters change budgets. A
failedorvoidedoutcome gives limit usage back only when reported with an approver token or by a signed provider event. Agent reports are recorded and change nothing, so an agent can't free its own budget. - Capabilities reveal thresholds to agents (on purpose, so they can plan). Turn it off with
"introspection": falsein the policy. - Callbacks are signed with your account's secret and go only to public HTTPS hosts. Verify the signature before resuming a workflow.
Independent verification (don't trust us)
| What | How you check it without trusting Auditail |
|---|---|
| A payout was approved | Signed receipt: Ed25519, bound to your account, the action and a SHA-256 hash of the payload, with an expiry (15 minutes for approvals, 5 for policy decisions). Verify offline with the open-source SDK. |
| A receipt is used once | redeemReceipt(), or your own record of claims.aid. |
| The audit log wasn't rewritten or cut short | Hash-chained log, plus signed checkpoints you keep, pushed automatically to your own write-once storage if you like (anchoring). Exports verify offline against them. |
| The examiner report wasn't edited | Signed by Auditail; npx -p auditail-sdk auditail-verify report.json. |
| What the code does | The SDKs and the server are MIT-licensed on GitHub. |
Public keys: /.well-known/auditail-keys.json, also pinned in the repository as AUDITAIL_PUBLIC_KEYS.json.
No lock-in
- Full export of your log at any time (
GET /v1/export), in a documented format that verifies offline. - Self-hosting: the server runs in your own Cloudflare account (guide).
Subprocessors
| Subprocessor | What for | Customer data? |
|---|---|---|
| Cloudflare, Inc. | Hosting, database, website | Yes: everything Auditail stores |
| Formspree | The waitlist form on auditail.io | Only what you type in the form |
| Google Workspace | Email with design partners | Only what you email us |
What we don't have yet
We'd rather you hear this from us:
- No SOC2 report or ISO 27001 certificate of our own yet. Cloudflare, our host, has both (Cloudflare Trust Hub), but that doesn't cover Auditail's own controls.
- No independent penetration test yet.
- A data processing agreement (DPA) is being prepared. Ask us if you need one before you start.
- No uptime SLA during the beta.
- Hosted data is in the APAC region only.
Reporting a vulnerability
Email [email protected] with "Security" in the subject. Please don't open a public GitHub issue. We'll reply within 2 business days. See SECURITY.md.