The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: This skill is specifically designed to handle and summarize untrusted external data (such as web search results or knowledge base documents). While the skill's primary function is to provide a 'Closed World' boundary and citation system to mitigate hallucinations and injections, the ingestion of external content remains a standard attack surface for indirect prompt injection. The skill provides solid boundary markers (e.g., [A:L12] citation tags) and instructions to ignore any external content not strictly grounded in the provided sources.
[COMMAND_EXECUTION]: The skill utilizes a local Python script scripts/ground_check.py to perform automated auditing of the generated text. This script performs benign string processing, regular expression matching, and file reading (read-only on source documents) to verify content against a whitelist. The instructions guide the agent or user to execute this script in the local environment.
[SAFE]: No malicious patterns, obfuscation, data exfiltration, or credential harvesting were detected. The Python scripts provided (ground_check.py and test_ground_check.py) use standard libraries and do not perform network operations. The file scripts/common_chars.txt is a legitimate dictionary of common Chinese characters used for frequency analysis and entity detection.

you-ju