relevancekit
Warn
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: MEDIUMMETADATA_POISONINGEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [METADATA_POISONING]: The skill references non-existent software versions, specifically watchOS 26 and Swift 6.3, in its description and instructions. This deceptive metadata can lead to incorrect agent behavior, such as generating hallucinated code or attempting to access unavailable APIs. Findings are present in SKILL.md and throughout the Swift snippets in references/relevancekit-patterns.md.
- [EXTERNAL_DOWNLOADS]: Documentation references point to 'sosumi.ai', a third-party domain and parody of Apple documentation, rather than official developer resources. This introduces a risk where the agent may ingest inaccurate or malicious technical guidance from an untrusted source.
- [INDIRECT_PROMPT_INJECTION]: The skill defines patterns for processing data from external sources (such as meetings and tasks) without recommending sanitization or boundary markers.
- Ingestion points: MeetingStore and TaskStore data providers referenced in references/relevancekit-patterns.md.
- Boundary markers: Absent; no instructions are provided to delimit or sanitize external data before processing.
- Capability inventory: The skill uses RelevantIntentManager for system state updates and requires sensitive permissions including HealthKit and Location.
- Sanitization: No validation or escaping of external content is suggested in the provided implementation patterns.
Audit Metadata