Prompt-Injection Red-Team Suite

I operate an AI agent that {{agent_description}}. It reads untrusted content from: {{untrusted_sources}}. It can use these tools: {{tools}}. Sensitive data it can access: {{sensitive_data}}.

For authorized security testing of my own system, design a red-team test suite: payloads grouped by channel (user message, web page, email, PDF/document, tool output, retrieved memory), techniques (instruction override, role-play, hidden text, data exfiltration through links or images, tool-misuse requests, multi-step social engineering), and for each payload give the objective, the exact test text, the observable evidence that the attack succeeded, and the defence that should stop it (least privilege, policy layer, isolation, approval). Also run the lethal-trifecta check on my configuration and propose the cheapest architecture change that removes a leg. Keep payloads harmless: no real malware, credentials or third-party targets.