The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external ad networks and universal links, creating a surface where malicious payloads could attempt to influence the agent's behavior.
Ingestion points: Processes JWS (JSON Web Signature) payloads using AppImpression(compactJWS:) in both SKILL.md and references/adattributionkit-patterns.md. It also extracts parameters from universal link URLs in the handleUniversalLink and handleReengagementURL functions.
Boundary markers: The instructions do not specify any delimiters or safety prompts for the agent to ignore or isolate instructions that might be embedded within the processed JWS strings or URL parameters.
Capability inventory: The skill is capable of updating conversion values, presenting StoreKit overlays, and opening universal links within the advertised app.
Sanitization: The implementation includes robust signature verification using CryptoKit to validate the integrity and origin of the JWS data before processing.
[METADATA_POISONING]: The References section of the skill contains links to https://sosumi.ai/documentation/* while labeling them as official Apple documentation. Although the domain is a known parody site in the tech community, pointing to non-authoritative sources for security-sensitive API implementations is potentially misleading to users and developers.

adattributionkit