Skip to main content

For Evaluators

This page is for teams deciding whether to trust Auditail with controls on AI agents that move money. We'd rather tell you where we stand than have you guess.

Where Auditail is today​

  • Stage: private beta. We launched in November 2026 and are onboarding our first design partners in payments, lending and neobanking.
  • Release pace: frequent, deliberate releases. Each one is listed in the changelog with what it adds or fixes.
  • What's covered by tests: the policy engine, approvals (including two-person approval), receipts, the audit log, the kill switch, reports, act-only keys and the provider integrations. The full suite (unit, API and SDK end-to-end) runs with one command: scripts/test-all.sh in the repository.
  • What's beta: the payment provider integrations follow each provider's published documentation and are tested with correctly signed requests. They haven't yet been verified against every provider's live sandbox.

Why it's low-risk to evaluate​

  • You don't have to trust our servers. The SDKs and the server are MIT-licensed. You can self-host in your own Cloudflare account, signed with your own key.
  • You don't have to trust our word. Receipts, checkpoints, exports and the examiner report are Ed25519-signed and verify offline with open-source tools.
  • No payment credentials. Integrations work through your provider's own authorization and approval steps.
  • Fails closed. If Auditail is unreachable, policy-checked actions are denied. Nothing runs unchecked.
  • No lock-in. Export your full log at any time in a documented, verifiable format.

Check our claims yourself​

ClaimHow to check it
Agents can't change their policy or approve themselvesRoute checks in worker/src/index.ts; tests in worker/test/policy.test.mjs and approvals.test.mjs
Splitting a payment doesn't get around a limitAtomic reservation in worker/src/policy.ts (reserve); concurrency test in policy.test.mjs
A receipt can't be reused or applied to another payloadverifyReceipt in sdk-js/src/verify.ts; redemption in worker/src/assurance.ts; tests in controls2.test.mjs
Rewriting or cutting short the log is detectableverifyEvidence with saved checkpoints; tests in evidence.test.mjs, which really edit the database
It fails closedpolicy_unavailable handling in the SDKs; tests in sdk-js/test/e2e.test.cjs and sdk-python/tests/test_e2e.py

What we don't have yet​

No SOC2 report or penetration test of our own, no uptime SLA during the beta, hosted data in one region (APAC), no SSO for approvers, and a DPA that's still being prepared. Details and workarounds are on the security page. Tell us which of these blocks you, because that sets our order of work.

A safe 30-day pilot​

  1. Week 1: connect one agent in shadow mode, against a sandbox payment account. Nothing is blocked; you see what would have been.
  2. Week 2: write the policy you'd want in production, and backtest it on the week's traffic.
  3. Week 3: enforce it in the sandbox and try to break it. Split payments, replay or tamper with a receipt, pay a new beneficiary, approve your own request, take Auditail offline mid-payout, and pull the kill switch.
  4. Week 4: export the evidence, verify it offline, and generate the examiner report for your risk team.

Prefer to keep everything in your own infrastructure? Run the whole pilot on a self-hosted server.

Contact​

[email protected]. We'll answer security questionnaires and walk your team through the design.