The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process third-party web content, creating an attack surface where malicious data on monitored pages could be interpreted as instructions.
Ingestion points: External content is retrieved using the requests library in SKILL.md within the PageMonitor.check_page and RSSGenerator.add_from_page methods.
Boundary markers: The skill includes a dedicated 'Untrusted content boundary' section that explicitly instructs the agent to treat retrieved material as data rather than instructions and recommends the use of <EXTERNAL_DATA> delimiters.
Capability inventory: The skill scripts have capabilities for network access (requests), file system writes (pathlib), and executing specific CLI tools like twarc2 using subprocess.run.
Sanitization: The instructions mandate URL validation, schema validation for extraction, and prohibit external content from authorizing command execution or credential usage.

page-monitoring