The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (booking confirmations, emails, tickets, PDFs, screenshots, and manual notes) to generate structured itineraries, creating a surface for potential indirect prompt injection attacks where instructions could be hidden in travel documents. * Ingestion points: Identified in SKILL.md Core Workflow (Step 1). * Boundary markers: The instructions lack explicit delimiters or warning instructions for the agent to isolate untrusted input data. * Capability inventory: The skill can write files and execute a local helper script for formatting. * Sanitization: PII redaction rules are present, but there is no mechanism to prevent the interpretation of instructions embedded within the source documents.
[COMMAND_EXECUTION]: The skill invokes a local script, scripts/itinerary_utils.py, for deterministic formatting and validation chores. * The script uses the Python standard library and implements input sanitization for generating map URLs, representing a controlled and safe use of external execution for skill-related tasks.

travel-itinerary