testing-prompt-injection-in-rag-pipelines

The skill is internally consistent and uses mostly official tooling/install paths, so it is not malware and does not show deceptive credential routing. However, it is a high-risk offensive security skill for AI agents because it enables prompt-injection, retrieval-poisoning, and document-exfiltration testing against live RAG targets.

This module is not indicative of covert malware, but it is purpose-built for offensive security testing of RAG/LLM systems (prompt-injection and embedding-retrieval manipulation). Its highest security risk is abuse potential and data leakage: it actively sends crafted instruction-subversion payloads to a user-supplied endpoint and can persist the returned assistant content (potentially including system prompts or secrets) into an on-disk JSON report. In a supply-chain/sandbox context, it should be treated as high-risk for misuse and carefully controlled (restricted execution, controlled network egress, and safe handling of outputs).