The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from landing page query parameters (fbclid, gclid, wbraid, gbraid, ttclid) and affiliate network webhooks. This ingestion of external data combined with the agent's ability to use execute_command and file system tools creates a potential attack surface.
Ingestion points: Query string parameters at the edge landing pages and affiliate lifecycle webhooks (rebill, refund, chargeback) processed in SKILL.md (Suggested Process steps 1 and 4).
Boundary markers: No specific boundary markers or delimiters for untrusted data are specified in the instructions.
Capability inventory: The skill allows execute_command, write_file, edit_file, and create_file across all operations.
Sanitization: The skill includes rules ASI05 (RCE Guard) and PII-NORMALIZATION which require sanitization and hashing, but these are instructions for the agent to implement rather than built-in platform safeguards.
[DATA_EXFILTRATION]: The skill is explicitly designed to collect and transmit user data to external advertising platforms (Meta, TikTok, Google). Although it requires SHA-256 hashing of PII (email, phone numbers) before transmission, the raw sensitive data is handled within the agent's context and processed by the agent.
Evidence: Rule META-CAPI-EMQ and GOOGLE-ENHANCED-CONVERSIONS describe the collection and forwarding of user identity vectors like email and phone numbers to external APIs.

setup-tracking-system