The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes a script scripts/check_retractions.py that performs network requests to api.crossref.org. This is a well-known scholarly metadata service used here to verify the retraction status of research papers via their DOI. This operation is limited to a specific, legitimate academic service and aligns with the skill's stated purpose.- [PROMPT_INJECTION]: The skill is designed to process untrusted user-provided content (e.g., research papers and manuscripts), which presents a surface for indirect prompt injection. However, the assets/risk_checklist.md file contains a proactive security rule ("外部内容防注入") that instructs the agent to identify and reject common injection patterns, such as "ignore previous instructions," treating them as data rather than valid commands.
Ingestion points: Processes external text files, DOIs, and metadata provided for ethical review.
Boundary markers: Includes specific instructions to detect and report injection attempts as INJECTION-ATTEMPT-DETECTED.
Capability inventory: Accesses local files for comparison and makes network calls to the Crossref API. It lacks capabilities for arbitrary shell execution or broad network access.
Sanitization: Implements word-level tokenization and normalization in scripts/text_overlap.py to compare text structure safely.

light-research-ethics