---
title: "Security at AgentTwin"
description: "How AgentTwin protects your code, accounts and test data: what runs where, who has access, what is kept and what is not certified yet."
canonical: "https://agenttwin.si/security/"
---

# Security at AgentTwin

This page describes the controls that are implemented today, as reviewed on November 6, 2026. It is not a certification. AgentTwin does not hold SOC 2 or ISO 27001 yet; we say so plainly so your review starts from facts.

## Your code stays with you

- Discovery and the main test loop (`npx agenttwin`) run on your machine. Syncing sends service names, workflow metadata, Git metadata and run receipts to your workspace. It does not upload application source or credentials.
- Tests run against digital twins with synthetic data. Calls a twin does not support are refused, never forwarded to the real service.
- Optional cloud runs upload only the workflow files, fixtures and configuration you explicitly select, into a short-lived E2B sandbox that is terminated after the run.

## Access control

- Sign-in with email, Google, GitHub, passkeys or your company's SSO (OIDC). SCIM 2.0 provisioning and deprovisioning for Okta and Microsoft Entra ID.
- Workspace roles: admin, member and viewer. Viewers cannot change anything; admin-only actions are enforced by the API, not the browser.
- Organization tokens for CI and SCIM, scoped to one workspace and a role, revocable at any time. OAuth access for third-party tools is scoped to one project.
- Every change in a shared workspace is recorded in the workspace audit log (who, what, when).

## Data handling

- Tokens and passwords are stored only as hashes. The CLI keeps its login in `~/.agenttwin/credentials.json` with owner-only file permissions, outside your repository.
- Workspace data is stored in PostgreSQL on servers operated by Contabo GmbH in Germany, with nightly backups kept for 14 days.
- Run receipts are kept for 30 days, up to 1,000 per account. Receipts can contain your application's output; use synthetic fixtures.
- The API enforces an origin allowlist for the dashboard, request size limits and rate limits.

## What is not done yet

- No SOC 2 or ISO 27001 report, and no completed third-party penetration test.
- Running inside your own cloud (BYOC / VPC) is not available.
- The workspace audit log is not yet exportable as a signed, tamper-evident file. Local runs keep a hash-chained audit trail (`npx agenttwin audit`).
- There is no self-service account erasure; [contact us](/contact/) and we delete it.

## Report a vulnerability

Please [contact us privately](/contact/) with the details. Do not open a public issue for a security problem. We acknowledge reports and keep you updated until they are fixed.

See also: [subprocessors](/subprocessors/), [privacy](/privacy/), [data processing agreement](/dpa/) and [safety boundaries](/docs/safety/).