The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file named reference-finder_audit_result_v1.json which contains a simulated security audit report. This file makes multiple unsupported claims about the skill's safety, status, and passing of security gates. This is a deceptive practice intended to mislead automated and manual security reviews.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the processing of untrusted scientific text and external API data.
Ingestion points: The skill accepts untrusted scientific text via command-line arguments or standard input in scripts/find_refs.py. The deceptive audit file reference-finder_audit_result_v1.json also acts as an ingestion point during the analysis process.
Boundary markers: There are no delimiters or instructions used to prevent the agent from following commands that might be embedded in the scientific text or the results fetched from PubMed.
Capability inventory: The skill possesses network access capabilities via urllib.request to the PubMed API.
Sanitization: The skill lacks sanitization for the text it processes, directly using input to construct API queries and displaying results without escaping potential control characters.
[DYNAMIC_EXECUTION]: In scripts/find_refs.py, the fetch_details function uses xml.etree.ElementTree.fromstring() to parse data returned from the PubMed API. As documented in the Python standard library, this module is not secure against maliciously constructed XML data, posing a risk of XML external entity (XXE) or denial-of-service attacks if the external data source were compromised or manipulated.
[EXTERNAL_DOWNLOADS]: The skill documentation and implementation show network requests to eutils.ncbi.nlm.nih.gov. These are standard API calls to the official PubMed service required for the skill's primary functionality.
[COMMAND_EXECUTION]: The skill provides instructions for installing dependencies using pip install requests and executing the script via the command line, which are typical development environment actions.

reference-finder