Skip to content
Argentum Field
Sign inDescribe a problem
Menu
EvidenceFor providersFor AI agentsPricingSecurity
Sign inDescribe a problem

Security and AI

Your text, the AI, and what protects them.

What happens to what you type, how AI is used and where it stops, and the controls behind the site. Nothing here is a badge.

Your text

What happens to what you type.

Your problem text is sent to the model service to produce the orientation. We do not store it as a case unless you are signed in and save one. Do not include personal, confidential or export-controlled information.

PROCESSINGOUR STORAGEyour textscreeningmodelorientationkept in this tab onlyforbidden tokens · scopestructured requestdirection + open questions→→→nothing on our side — not written to our databaseonly if a signed-in workspaceSAVES A PRIVATE CASE
FIG. 2Four steps. Your draft and the current answer stay in this browser tab so that leaving the form does not lose them; that never reaches us. Our own storage stays empty unless a signed-in workspace saves a private case. The model service processes the text to produce the answer; the privacy notice names what it retains and the short-lived handoff a workspace can choose.
hostingRailway · isolated application and PostgreSQL servicesmodel serviceOpenAI API · bounded orientation laneregion of recordunknown

Argentum Field uses the OpenAI API. API data is not used to train OpenAI models by default. Retention and processing details are described in the privacy notice; transactional email and payment providers appear there only where those features are enabled.

How AI is used

What the model may do, and what it may not.

Field Agent results are shown without prior human review. Anything routed to a provider or commercially committed requires human approval, and that approval and its date travel with the document.

It may

  • Interpret a problem written in plain language and map it onto a controlled method registry.
  • Explain retrieved, labelled evidence.
  • Draft text that a person then reviews.

It may not

  • Produce or restate a price, or add a fact or raise the level of one.
  • Confirm availability, capacity or feasibility.
  • Select a provider for routing, commit anyone to anything, or make a legal, export-control or medical determination.
Disclosure. The Field Agent provides AI-assisted orientation from Argentum Field. It may be incomplete and is not a scientific conclusion, provider confirmation or commercial offer. That notice appears under the problem field and on every result. Deterministic screening runs before any model call and again before routing; a flagged request stops for a person, and nothing on this site is export-control, state-aid or legal advice.

Report a vulnerability

Tell us before you tell anyone else.

Write to info@argentumfield.com with the steps to reproduce, and give us a chance to fix it before you publish.

Report a vulnerability
Security controls in detail

Implemented in the application

Each of these is written and tested. Whether a given deployment has it switched on is a separate question, answered next.

  • Single-use sign-in links and session bearers stored only as purpose-separated hashes, with an explicit confirmation step so a link scanner cannot consume a link.
  • A same-origin check on every mutation made from a session cookie.
  • A second factor for platform administrators and provider owners, with an encrypted authenticator secret, bounded attempts and one-use hashed recovery codes.
  • Server-side role and tenant checks, with negative isolation tests.
  • Per-address, per-network and per-subnet sign-in quotas, stored as keyed hashes rather than addresses.
  • Atomic usage quotas and a deployment-wide model-spend ceiling, with a per-call reservation.
  • Separate kill switches for the public shell, the catalogue, provider discovery and orientation.
  • Payment pages hosted by the payment provider. No card or bank detail is stored here, and a checkout alone never grants a plan; a signature-verified webhook does.
  • Machine credentials that are scoped, expiring and revocable, with idempotency keys and per-credential rate policies.

Confirmed in this deployment

  • A nonce-based strict Content-Security-Policy and the hardening headers, on every response.
  • No third-party script, style, font or image on any public page. Nothing on this site calls another origin.
  • A database runtime role that is not the owner, cannot change the schema, and can read the research evidence only through two sanitised projections. It is verified when the container boots.
  • Continuous integration running typecheck, tests, secret scanning, a dependency advisory gate, pinned actions and a non-root container.

Not yet

  • No external penetration test has been completed.
  • No ISO 27001 or SOC 2 certification is claimed, because none has been obtained.
  • Current staging uses one application replica. Multi-replica and failover behaviour has not been exercised.
  • Edge throttling, an incident-response exercise and backup-restore evidence are pre-launch gates, not finished work.
  • Customer result files are not open to real data until private storage, malware scanning and an agreed retention schedule are in place.