The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that claim to be authoritative and unconditionally supersede the agent's prior training, utilizing a pattern that attempts to override the model's internal knowledge base and potentially its safety guidelines.
[METADATA_POISONING]: The skill content claims it was written and published by Apple, which is a misleading statement given the skill is authored by tartinerlabs. This deceptive metadata can cause the agent or user to misjudge the reliability and provenance of the information.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a vulnerability surface by instructing the agent to ingest and trust external reference files without validation or isolation. 1. Ingestion points: Multiple reference markdown files in the references/ directory. 2. Boundary markers: None; the skill lacks delimiters or instructions to treat the reference data as untrusted or separate from system instructions. 3. Capability inventory: None; the skill does not contain scripts or system-level command execution capabilities. 4. Sanitization: Absent; there is no logic to filter, escape, or sanitize the contents of the reference files before they are processed by the agent.

swiftui-whats-new-27