prompt-injection-testerlisted

# Prompt Injection Tester ## Purpose / Amaç Drive a repeatable, evidence-based prompt-injection assessment against a target LLM application and produce a scored report. The skill covers **direct** injection (attacker text arrives in the user turn) and **indirect** injection (attacker text arrives through retrieved documents, tool output, web pages, file contents, or email the model ingests). It is bilingual by design: Turkish-language payloads routinely bypass English-only guardrails, so every family is tested in both English and Turkish. The output is not "the model said something bad." It is: **which family bypassed which control, with which evidence, at what severity, mapped to OWASP LLM01 and ATLAS** — plus an over-refusal measurement so a defense that simply blocks everything does not score as "secure." ## When to use Use this skill when the task is any of: - "Test / audit / pentest this chatbot (or agent, or RAG system, or `/chat` endpoint) for prompt injection." - "Can someone leak our system prompt / exfiltrate the canary / override our instructions?" - "We added a guardrail — prove whether it actually blocks injection without over-blocking." - "Check our tool-using agent for excessive agency triggered by injected instructions." Do **not** use this skill to attack a system the operator has not authorized (see Limits & ethics). ## Prerequisites & scope (do this first) 1. **Confirm authorization.** You must have explicit permission to test the target. Record the