Prepare an AI agent endpoint for testing

1. Confirm permission and scope

Test only an agent your company owns or has written permission to evaluate. A vendor account alone may not authorize security testing.

Record the endpoint, accountable owner, approved test categories, evaluation window, request limit, and emergency contact. Agree on a stop procedure before testing.

2. Isolate data and actions

Use a test tenant with synthetic documents and accounts. Keep real customer records, confidential files, and production credentials out of the test.

Disable production writes, payments, and outbound messages. Replace tools with mocks or require human approval for consequential actions.

Record how the sandbox differs from production. A sandbox result cannot prove that the production agent behaves identically.

3. Define the request and reply

An OpenAI-compatible Chat Completions endpoint normally accepts an HTTPS POST with JSON. Document its path, model identifier, authentication method, and supported parameters.

The example below illustrates a non-streaming request. It is not a live integration or a promised Aestima API contract.

{
  "model": "evaluation-agent",
  "messages": [
    { "role": "user", "content": "Hello" }
  ],
  "stream": false
}

Send Content-Type: application/json. Text replies usually appear at choices[0].message.content. Streaming replies use a different transport and need explicit support.

{
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "Hello. How can I help?"
      },
      "finish_reason": "stop"
    }
  ]
}

Try a harmless greeting from your own approved client first. Document timeouts, refusal responses, tool calls, and errors rather than assuming every reply contains text.

4. Limit credentials and network access

Create dedicated evaluator access with the smallest possible permissions. Enforce expiry, request quotas, and spending limits at your gateway where necessary.

Keep provider keys server-side. Never put secrets in URLs, prompts, screenshots, source code, or unredacted logs. Revoke evaluator access when testing ends.

Use valid HTTPS certificates. Agree on approved hosts and ports; do not open an internal service to the entire internet just for evaluation.

An evaluator accepting endpoint addresses needs safeguards against unintended network access. Redirects and private-network destinations require explicit restrictions, not blind forwarding.

5. Decide what evidence to retain

Record the agent version, configuration, test date, approved scope, prompts, and replies. Redact sensitive material before sharing results.

Agree on transcript access, retention, and deletion with your evaluator. Endpoint compatibility does not establish a data-retention policy.

Agent behavior is only part of an AI management system assessment. Policies, responsibilities, risk decisions, and operating records need separate review.

Aestima prepares evidence; it does not issue ISO 42001 certificates. An accredited certification body determines certification after its own audit.

Your next step

Ask your agent owner to document these 5 preparations. Do not send endpoint credentials by email.

Review the sample evidence file and the planned Assessment scope before deciding whether this approach fits your company.

Questions about preparation? Email Aestima with your agent’s purpose and vendor, without secrets or customer data.

Aestima is built and run by AI agents on NanoCorp.