Security at mtge.ai
mtge.ai is a preview. This page describes the security controls that exist in it today and the ones that do not yet. None of them has been independently audited or tested by a third party.
Hosting and data protection
The preview's web application, API and database run on a single machine that we operate, not on audited cloud infrastructure. Language-model requests go either to a model on a second machine we operate, over an encrypted private network, or to a third-party model provider's API. Only the web application is published to the internet, through a Cloudflare Tunnel; the API listens only on its own machine and is reached through the web application's sign-in gate. Traffic is encrypted in transit between your browser and Cloudflare and through the tunnel. We do not claim encryption at rest for the preview. One person holds administrative access to these machines and the application.
Identity and access
Product screens and the API sit behind a sign-in gate: a preview credential, checked on the server, that issues a signed session cookie lasting up to seven days. Single sign-on (SAML or OIDC) and multi-factor authentication are not available. The application checks a role-based capability for every action, but in the preview every signed-in visitor acts as the same identity, so roles do not yet separate individual users.
Source integrity and provenance
Every monitored source records its publisher, and every version of it records a content hash and, where the source states one, an effective date. Answers cite the exact passage they rest on, and the platform will not answer from a source whose rights are unresolved or from a version that has not been published.
Append-only audit trail
The audit log records which identity asked what, the sources an answer cited, and the model and prompt version that drafted it, alongside other actions taken in the product. Entries are hash-chained, and a database trigger refuses updates and deletes. A nightly job is scheduled to verify the whole chain and record the result. The chain makes tampering detectable, not impossible: someone with full control of the database could rewrite it, because its head is not yet anchored anywhere outside that database.
AI safeguards
Generated answers are grounded in retrieved source text and returned with citations so reviewers can verify them. When the retrieved sources do not support an answer, the question is escalated rather than answered. We do not use Customer Content or user questions to train models.
Operations and resilience
Automated checks (tests, lint, type checks, a build, end-to-end tests, migration checks, and a Python dependency audit) run on every change to the main branch, but nothing blocks a change that fails them: the one administrator pushes changes directly rather than through pull requests, and no second person reviews them. The preview is deployed by hand, without a staging environment. It has no committed recovery point or recovery time objective, and no tested, encrypted, off-site backup.
Compliance status
mtge.ai has not undergone a SOC 2 examination or any other independent audit, and no independent penetration test has been performed. Its controls have not been formally mapped to the SOC 2 Trust Services Criteria or to the GLBA Safeguards Rule. There are no audit reports to share.
Responsible disclosure
We welcome reports from security researchers. Please give us reasonable time to remediate before public disclosure; we will not pursue action against good-faith research that respects customer data.