The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: the skill includes a client-side utility for capturing marketing touchpoints that ingests untrusted input from the browser's location string without sanitization.\n
Ingestion points: The captureUTMTouchpoint function in SKILL.md reads parameters directly from window.location.search (specifically utm_source, utm_medium, utm_campaign, gclid, and fbclid).\n
Boundary markers: The implementation does not include any delimiters or instructions to ignore embedded prompts within the captured data.\n
Capability inventory: Touchpoints are stored in localStorage and subsequently transmitted to a server-side endpoint during order creation (createOrder).\n
Sanitization: The code lacks validation or escaping for the string values extracted from the URL before they are stored and transmitted.\n
Risk: An attacker could craft a URL containing malicious instructions in the UTM parameters. If these touchpoints are later retrieved and processed by an AI agent (e.g., for generating a marketing report), the agent may inadvertently follow the injected instructions.

attribution-modeling