The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local Python script push_to_zotero.py via the command line to process citation data and communicate with the Zotero desktop application.
[DATA_EXFILTRATION]: The Python script push_to_zotero.py implements a PDF download feature (download_pdf) that accepts browser cookies and a target URL. There is a risk that ScienceDirect session cookies could be exfiltrated if the pdfUrl provided in the JSON input points to an attacker-controlled domain rather than the expected ScienceDirect infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill ingests article metadata (titles, abstracts, and PIIs) from ScienceDirect search results. This data is then processed and sent to the Zotero API, creating a surface where malicious instructions embedded in bibliographic records could potentially influence agent behavior.
Ingestion points: ScienceDirect search result pages and metadata fields.
Boundary markers: None identified in the script or prompt interpolation.
Capability inventory: JavaScript execution in the browser (evaluate_script), network requests to ScienceDirect and local Zotero APIs, and file system read access for RIS/JSON input files.
Sanitization: The script parses JSON and RIS data but does not perform domain validation on the pdfUrl before sending cookies.
[DYNAMIC_EXECUTION]: The skill uses evaluate_script to interact with the ScienceDirect DOM, extract session-specific security tokens (t), and submit form data. This is standard behavior for web-based automation tasks.

sd-export