The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill contains a fetch_title function in scripts/source_db.py that retrieves and parses HTML <title> tags from remote URLs. If a researcher adds a URL to a malicious website that contains prompt injection instructions within its title tag, these instructions will be stored in the citation database. When the agent later searches the database or generates a bibliography, it may ingest these instructions, leading to unintended behavior.
Ingestion points: fetch_title in scripts/source_db.py fetches data from remote HTTP/HTTPS URLs via urllib.request.urlopen.
Boundary markers: The skill does not employ specific delimiters or instruction-bypass warnings when presenting stored titles to the agent during search or export.
Capability inventory: The skill allows the agent to write files to the local file system (export) and make outbound network requests (title fetching and health checks).
Sanitization: While the script uses a TitleParser to extract the title text, it does not sanitize the resulting string for potentially malicious instructions.
[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands to run Python scripts (scripts/source_db.py and scripts/url_health.py) to manage the citation database and perform network-based health checks.

source-tracker