The Agent Skills Directory

[METADATA_POISONING]: The documentation links provided in the reference material point to the non-official domain sosumi.ai rather than the official Apple developer portal (developer.apple.com). This represents deceptive metadata that could lead users or agents to untrusted third-party content.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data via Apple Push Notification service (APNs) payloads, which creates an indirect prompt injection surface.
Ingestion points: External JSON payloads containing content-state data are ingested to update Live Activity views (e.g., in references/activitykit-patterns.md).
Boundary markers: The instructions do not define delimiters or defensive instructions to ensure the agent ignores potentially malicious commands embedded within the dynamic content-state fields.
Capability inventory: The skill uses Activity.update and ActivityConfiguration to modify system-level UI components (Lock Screen, Dynamic Island) based on the ingested data.
Sanitization: The skill recommends keeping sensitive data in the app but provides no specific input validation, escaping, or schema enforcement for the external data processed at runtime.
[METADATA_POISONING]: The skill metadata claims to target future platform versions (iOS 26 and Swift 6.3), which is misleading and potentially deceptive information regarding current platform capabilities.

activitykit