The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill interacts exclusively with official and trusted Taiwanese government domains, including law.moj.gov.tw, judicial.gov.tw, and mojlaw.moj.gov.tw. These sources are well-known services for legal documentation and do not represent a data exfiltration risk. The skill does not access sensitive local files or environment variables.
[COMMAND_EXECUTION]: The tests/test_grade_research.py script uses subprocess.run to call a local structural grader script (scripts/grade_research.py). This is a deterministic process used to validate the structure and content of generated legal memos against predefined JSON schemas. It does not facilitate arbitrary command execution or process untrusted remote scripts.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests external data from web searches, it mitigates potential injection risks through a 'plan-first' interaction model where the user must approve the search strategy. Additionally, the final outputs are subjected to a structural grader that enforces strict formatting rules and mandatory legal disclaimers.
[SAFE]: The skill demonstrates high security maturity by implementing versioned protocols, schema validation for its state and output files, and clear boundary markers for AI-generated content. It correctly implements localized legal constraints (Path A discipline) and mandatory escalation paths for high-stakes queries.

legal-research