superagent_

// security workers

Agents

[ view markdown ]

Agent Red Team runs perform automated black-box adversarial testing against a deployed AI agent. They discover capabilities, exercise relevant attack hypotheses, preserve evidence, and write verified security-boundary failures to a red-team report without requiring source-code access.

Create and configure an Agent

  1. Open Agents → Connected and select Add new.
  2. Enter a name and select an agent type.
  3. Optionally associate a connected repository.
  4. For Coding Agents, copy the one-time Runtime Guardrails pairing command.
  5. Open the Agent detail page and enable Red teaming.
  6. Configure the application or API target, credentials, headers, throttle, schedule, and optional specific prompt.
  7. Select New run, or allow the daily, weekly, or monthly schedule to start it.

Each run is linked to the persistent Agent. Manual, scheduled, and API-triggered runs appear together on the Agent detail page and retain their full report detail.

Start an ad hoc agent run

  1. Open Red team and select New report, or go to /app/reports/new.
  2. Choose Agent.
  3. Select an agent type (chatbot, voice agent, coding agent, AI workflow, data pipeline, or other).
  4. Choose the target type: Application or API.
  5. Enter the public target URL, configure optional encrypted headers, and add a specific prompt when needed.
  6. Submit to provision the isolated runner and open the live detail page.

The report name defaults to the target host. Agent reports created in the dashboard are labeled Manual in the report list. Reports created through REST or MCP are labeled API.

Application targets

Use Application when the AI agent is accessed through a browser interface. Superagent generates test credentials that you can view, edit, or copy before starting the report. You can also add encrypted headers for staging gates, tenant selection, or authentication.

The isolated runner interacts only with the configured origin. Browser interaction evidence can appear in the Evidence tab when recording is available.

API targets

Use API when the AI agent is exposed through an HTTP or WebSocket endpoint. Enter only the endpoint URL. Add authentication or tenant headers under Headers.

Describe the API contract in Specific prompt, including:

  • HTTP method and example JSON payload
  • Where the attack prompt belongs
  • Required model or deployment identifier
  • Expected response or streaming shape

For example:

POST JSON like:
{
  "model": "qwen/qwen3.7-flash",
  "stream": false,
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "<attack prompt>" }
  ]
}

API reports expose Logs and Findings. They do not show browser evidence.

REST API and MCP

Programmatic clients can create and configure persistent Agents with:

  • POST /api/v1/agents and PATCH /api/v1/agents/{agent_id}
  • MCP tools create_agent and update_agent

Use the Agent endpoints for reusable identity and scheduled configuration. For an ad hoc run, use:

  • POST /api/v1/reports/agent
  • MCP tool create_agent_report

Both report operations accept target_type as web_app or api. See the Agents API and Reports API for request fields and examples.

The agent detail page

Each automated Agent report lives at /app/reports/agent/[id]. From there you can:

  • Follow expandable runner logs
  • Review findings and risk counts
  • Inspect the target, assignees, cost, and runtime status
  • Delete or rerun the report
  • Review browser evidence for application targets

Test agents safely

  • Use dedicated test credentials and scoped API keys
  • Rotate any key pasted into logs or chat
  • Describe API payloads precisely in the specific prompt
  • Apply a request throttle when testing shared or metered targets
  • Pair agent tests with repository reports when the agent depends on specific code or tools in a repo
  • Re-run tests after material model or policy changes

Next steps