The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides example curl commands for probing the reachability and security of common vector databases, including Chroma (port 8000), Qdrant (port 6333), and Weaviate (port 8080).
[PROMPT_INJECTION]: Instructions include example payloads designed to test for persistent corpus poisoning. These payloads use common injection markers like 'IMPORTANT INSTRUCTION FOR THE ASSISTANT' and instructions to suppress mentions of the behavior to test the target's vulnerability to hidden instructions.
[DATA_EXFILTRATION]: The skill describes methodologies for verifying data leakage using Out-of-Band (OOB) techniques. It provides a placeholder URL (https://OOB.example/leak) as a template for confirming that an injected instruction can successfully exfiltrate context from a target system.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and defines the attack surface for indirect prompt injection within RAG pipelines. It specifically targets the vector storage and retrieval layer, providing guidance on how untrusted document uploads can be used to influence an agent's behavior across different user sessions.
Ingestion points: The skill guides the agent to process external data sources (documents, vector database query results).
Boundary markers: Recommends the use of specific 'False-Positive Gates' and 'run-twice' rules to verify that behavior is caused by injection rather than confabulation.
Capability inventory: Utilizes network probing (curl) and instruction generation to interact with and test target RAG implementations.
Sanitization: The skill emphasizes the need for 'Second-session' validation and 'Verifiable cross-tenant artifacts' to ensure findings are accurate and not false positives.

hunt-rag-vector