The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill queries the Crossref API (api.crossref.org) to check the retraction status of academic publications via the check_retractions.py script. Crossref is a well-known and trusted service in the research community. The script uses a standard User-Agent and a "polite pool" mechanism via an optional email address provided in an environment variable, which is standard practice for this service.- [COMMAND_EXECUTION]: Includes Python scripts (check_retractions.py, stat_consistency.py, text_overlap.py) that perform local data processing. These scripts are self-contained, use only the Python standard library, and do not execute arbitrary shell commands or external binaries.- [DATA_EXFILTRATION]: Network activity is strictly limited to identifying retracted papers based on DOIs. There is no evidence of sensitive local file access or unauthorized data transmission to third-party servers.- [PROMPT_INJECTION]: The skill implements defensive measures against prompt injection. In assets/risk_checklist.md, it includes a specific check for "External content anti-injection" that instructs the agent to treat instructional-looking text in research data (e.g., "ignore previous instructions") as data rather than commands.- [INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface when processing untrusted research materials but provides adequate defensive guidelines.
Ingestion points: Research papers, manuscripts, and software project files provided for compliance review.
Boundary markers: Instructions explicitly warn the agent to detect and report injection attempts (like "ignore above instructions") found within the analyzed data.
Capability inventory: The skill utilizes network-based retraction lookups and local statistical/textual analysis scripts.
Sanitization: Defensive logic is embedded in the risk checklist to ensure instructions are not executed if found in user-provided research data.

light-research-ethics