The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes climate news headlines and disaster event descriptions from external feeds, creating an attack surface where malicious instructions could be embedded in data.
Ingestion points: Data retrieved from worldmonitor.app endpoints (SKILL.md).
Boundary markers: The skill contains a "Content safety" section that explicitly instructs the agent to disregard and not execute directive-like text found in responses.
Capability inventory: Uses curl for network retrieval and jq for JSON processing.
Sanitization: Relies on instructional mitigations rather than programmatic sanitization of the incoming data stream.
[EXTERNAL_DOWNLOADS]: The skill retrieves climate data and disaster event records from the WorldMonitor API infrastructure.
[COMMAND_EXECUTION]: Employs standard system utilities curl and jq to fetch and parse external disaster data.
[PROMPT_INJECTION]: A static detector flagged a potential prompt injection pattern in the instructions. However, manual review confirms this is a false positive; the text "ignore previous instructions" appears within a defensive context, providing safety guidelines for the agent rather than attempting to bypass safety filters.

track-climate-hazards