Field manual · AI for authorized security testing · Edition November 2026 · Updated 2026-11-18

Tool · AGPL-3.0

Shannon setup: a models.json provider and one model id

Shannon runs one model for the whole scan, described in a Pi models.json file passed with --models-config when the model is not in its catalogue.

Set SHANNON_AI_MODEL to <provider>:<model-id>, put the key in SHANNON_AI_API_KEY, and use host.docker.internal rather than localhost for a server on your machine.

Updated November 18, 2026Authorized work only
01Fit

When Shannon is the right agent

Shannon's README calls it "an autonomous AI pentester for web applications and APIs." (Shannon) The provider doc says "If a safeguard engages mid-run, the model can refuse, and the scan fails partway through rather than at the start." A model that does not follow its tool-use constraints reliably "will produce weaker pentests than a frontier model." (Shannon docs) Kimi K3.1 Pro is the pick.

02Setup

Describe the model, then name it

From Shannon's AI provider document, read November 18, 2026.

01

Install from the project

The documented commands are npx @keygraph/shannon and ./shannon. You need Docker and the source. Shannon repository.

02

Write models.json for an OpenAI-compatible API

The provider name is yours, openai-completions is the chat dialect, and documented base URLs end in /v1. Leave the key as YOUR_API_KEY. Custom model configuration.

models.jsonmodels.json
{
  "providers": {
    "silkcompute": {
      "baseUrl": "https://api.silkcompute.com/v1",
      "api": "openai-completions",
      "apiKey": "YOUR_API_KEY",
      "compat": { "supportsDeveloperRole": false, "maxTokensField": "max_tokens" },
      "models": [
        {
          "id": "kimi-k3.1-pro-uncensored",
          "name": "Kimi K3.1 Pro",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 1048576,
          "maxTokens": 131072,
          "cost": { "input": 3, "output": 8, "cacheRead": 0.3, "cacheWrite": 0 }
        },
        {
          "id": "deepseek-v4.1-flash-1017-uncensored",
          "name": "DeepSeek V4.1 Flash 1017",
          "reasoning": true,
          "input": ["text", "image"],
          "contextWindow": 1048576,
          "maxTokens": 131072,
          "cost": {
            "input": 0.24,
            "output": 0.9,
            "cacheRead": 0.038,
            "cacheWrite": 0
          }
        }
      ]
    }
  }
}

Prices per 1M tokens are Silk Compute list prices, read November 18, 2026. Get a key.

03

Name the model with a colon

The form is SHANNON_AI_MODEL=<provider>:<model-id>. The prefix must match the key inside providers.

Environmentshell
export SHANNON_AI_API_KEY="$SILK_API_KEY"
export SHANNON_AI_MODEL="silkcompute:kimi-k3.1-pro-uncensored"
npx @keygraph/shannon start -u https://app.example -r ./repo --models-config ./models.json

Replace https://app.example with the in-scope URL. The line above is silkcompute:kimi-k3.1-pro-uncensored. Swap in deepseek-v4.1-flash-1017-uncensored for volume work.

04

Check the model before a full scan

--validate-model runs the credential check and stops, with no report.

Preflightshell
npx @keygraph/shannon start -u https://app.example -r ./repo --models-config ./models.json --validate-model
Rules of engagement

Only on systems you own, a signed scope, a CTF or a lab. A custom model does not widen the scope.

"Use host.docker.internal, not localhost. The scan runs inside a container." (Shannon docs) If a reasoning model is rejected, the doc shows {"supportsDeveloperRole": false, "supportsReasoningEffort": false}. The file above sets maxTokensField to max_tokens.

03Prompt

Keep the run inside the repository you were given

Prompt

Authorized white-box test of the repository and URL in this run. List the access-control and input-handling questions a tester should answer. Stay inside SCOPE.md. Do not write exploit code.