Requirements and permissions

Check these requirements before you add your first target. The most important one is knowing exactly how your AI application is reached and how it authenticates requests.

Account requirements

• A QuilrAI tenant with the Red Teaming module enabled.

Know your target application

Quilr Red Teaming can only test what it can reach, authenticate to, and talk to in the format the target expects. The team that runs red teaming, usually security, rarely builds the AI applications, agents, and workflows it tests. The details you need sit with the development team that owns each one, so plan to work with them before you add a target.

For each application, agent, or AI workflow, get these five things from its development team:

• Access: the endpoint URL, HTTP method, and any network requirements such as VPN or IP allowlisting.
• Authentication: how the application verifies the caller, and a valid test credential.
• Authorization: the role, permissions, and data the test account is allowed to reach.
• Request format: the fields, types, and content type the endpoint accepts.
• Response format: where the application's answer, session ID, and actions appear in the reply.

Missing or incorrect details here are the most common cause of setup delays. Authorization matters as much as authentication: the test account's permissions define what the Red Team Agent can attempt, so a test account with too little access hides real risks, and one with too much access can cause real side effects. Agree on the test account's permissions with the development team.

There are two ways to gather this information. Start with the quick path, and move to the detailed checklist when the target needs it.

Quick path: share one working request

For most targets, one working example answers nearly every question. Ask the development team for a successful API call that includes the authentication method, required headers, request body, and expected response, in any of these forms:

• A working curl command
• A Postman collection
• OpenAPI (Swagger) documentation with a working request example

For example, this request to a document summarization API:

curl -X POST \
  https://documents.example.com/summarize \
  -H "X-API-Key: $API_KEY" \
  -F "file=@sample.pdf"

tells Quilr everything it needs to connect:

Component	Value
Endpoint	https://documents.example.com/summarize
HTTP method	POST
Authentication	API key
Authentication header	X-API-Key
Request format	multipart/form-data
Required input	file (an uploaded document)

warning

Share credentials only through an approved secure channel. Never put real keys or tokens in documentation, tickets, or screenshots. Use a placeholder such as $API_KEY, as in the example above.

Detailed checklist

Use this checklist for targets with custom authentication, short-lived tokens, file uploads, multi-step workflows, or tools and MCP integrations. Share it with the development team that owns the application, agent, or AI workflow, or walk through it with them in a short working session.

Area	Questions for the development team
Application access	How is the application reached programmatically? What is the exact endpoint URL and HTTP method? Does access require a VPN, IP allowlisting, private networking, or special TLS certificates?
Authentication	Which mechanism does the application use: API key, bearer token, OAuth 2.0, session cookie, Basic authentication, mutual TLS, or none? How will Quilr get valid test credentials, and how are they renewed when they expire?
Authorization	What role, permissions, and data access does the test account have? Is it separated from real user data?
Request format	Does the endpoint accept JSON, multipart/form-data, XML, binary payloads, or WebSocket messages? Which fields are required, and what are their names and types? For file uploads, what is the field name, which file types are supported, and what is the size limit?
Response and sessions	What does a successful response look like? Does the application need a session ID, conversation history, or several requests in sequence? For asynchronous workflows, how are results retrieved?

Response field names are specific to each application. For example, an application might return:

{
  "reply": "Here is the document summary...",
  "session_id": "session-123",
  "actions": []
}

Here you map reply as the response text and session_id as the session value in the target's request and response mapping. See Red team an agent.

note

The system prompt and tool list are what let the Red Team Agent tailor attacks to the target (a white-box test). Ask the development team for them early. If they cannot be shared, testing still runs as a black-box test, but attacks are less targeted.

Choose an approach

Target	Start with
Simple REST API with straightforward authentication	Quick path
The development team already has a working curl command or Postman collection	Quick path
Custom authentication or short-lived session tokens	Detailed checklist
Document uploads or other multimodal inputs	Detailed checklist
Stateful or multi-step workflows	Detailed checklist
Agents with tools or MCP integrations	Detailed checklist
White-box assessment	Detailed checklist

Readiness checks

Before you launch an assessment, confirm each of these with the development team:

Check	Passes when
Connectivity	Quilr reaches the application through the intended endpoint.
Authentication	The test credential is accepted, or anonymous access is confirmed.
Request mapping	Quilr can send every required field in the expected format.
Response mapping	Quilr extracts the application's answer from the response.
Testing authorization	The target, environment, permitted actions, and boundaries are approved.

tip

Send at least one harmless end-to-end request with Test connection before you launch an assessment. A successful network connection alone does not mean the application is ready to test.

danger

Get written authorization from the owner of every target before you test it, including third-party vendor AI. Testing a system you do not own or operate can breach its terms of service.