/guardx:red-team — Agentic Red Team Attack

Trigger

User invokes /guardx:red-team or asks to run an adversarial / red team test on a system prompt.

Steps

1. Ask the user: "Please provide the system prompt to attack — paste it directly or give me a file path."

2. If a file path is provided, use the Read tool to load the file contents.

3. Ask: "Which attack strategy would you like?

   • blitz — single pass, all attack categories, fast (~1 min)
   • thorough (default) — 3 adaptive phases, deeper analysis (~5 min)
   • stealth — slow, low-sophistication probes first, mimics a real patient attacker
   • goal-hijack — 5 objective manipulation techniques: objective substitution, priority inversion, scope expansion, authority override, consequentialist framing"

   If goal-hijack is selected, also ask: "What capability is the attacker trying to get the agent to perform? (e.g. 'exfiltrate user data', 'bypass access controls')"

4. Warn the user: "Red team scans take 3–8 minutes and use significant tokens. Proceed?"

5. Call the run_red_team MCP tool:

   Tool: run_red_team
   Arguments:
     systemPrompt: <the prompt text>
     strategy: <blitz | thorough | stealth>
   

6. When the result returns, automatically invoke /guardx:interpret on the result to present structured findings.

7. Offer to generate a report: "Generate a detailed report? Run /guardx:report or say 'generate HTML report'."

Notes

• Red teaming is more aggressive than a standard scan — it adapts strategy based on what works.
• The thorough strategy runs 3 phases: broad recon → targeted escalation → deep extraction.
• Results are auto-saved to scan history with a scanId for later retrieval.