The Agent Skills Directory

[METADATA_POISONING]: The skill references fictional software versions (iOS 26+ / iPadOS 26+) and provides links to an unofficial domain (sosumi.ai) as if it were the official Apple documentation source. This deceptive metadata could lead an agent or developer to rely on untrusted information.
[INDIRECT_PROMPT_INJECTION]: The skill defines patterns for processing external metadata which creates a surface for indirect prompt injection.
Ingestion points: label in WakeUpMetadata, and recipeName/stepDescription in CookingTimerMetadata within references/alarmkit-patterns.md.
Boundary markers: Absent; the skill does not instruct the agent to use delimiters or ignore embedded instructions within these data fields.
Capability inventory: The skill focuses on UI presentation and alarm scheduling via AlarmManager; no direct file-system writes, subprocess execution, or network exfiltration capabilities are implemented in the provided scripts.
Sanitization: Absent; no validation, escaping, or filtering is applied to the metadata strings before they are rendered in the system UI.

alarmkit