Prompt Injection Test Pack Builder

WHY THIS EXISTS

Built for AI-agent work that needs evidence, not vibes.

Generate a safe prompt-injection test pack for RAG and tool-using agents: attack cards, expected refusals, allowed actions, and JSONL eval rows. The useful output is a visible table plus a receipt that names input, checks, limits, and next action.

• Creates defensive eval cases without giving operational bypass guidance.
• Covers prompt injection, hidden instruction, tool misuse, and data exfiltration patterns.
• Exports JSONL rows that can be pasted into an agent evaluation harness.

Boundary: Not for bypassing real systems, stealing prompts, malware, credential theft, or offensive security instructions.