The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file named 'kv-design_audit_result_v1.json' that purports to be a formal security audit result. This file contains fabricated safety scores and assertions designed to mislead users or automated systems into believing the skill has already passed a rigorous review. Additionally, the 'Source' metadata in SKILL.md refers to a 'medical-research-skills' repository, which is inconsistent with the 'kv-design' functionality, suggesting potential metadata corruption or deceptive repurposing.
[INDIRECT_PROMPT_INJECTION]: The inclusion of the fake audit file constitutes a self-referential pattern (Category 8e) aimed at injecting a 'SAFE' verdict into the analyzer's environment.
Ingestion points: 'kv-design_audit_result_v1.json'
Boundary markers: Absent from the skill's instruction set.
Capability inventory: The skill includes 'scripts/generate_kv.py' which utilizes the 'requests' library for network operations and handles local file system writes.
Sanitization: No validation or sanitization is performed on the contents of the audit result file.
[COMMAND_EXECUTION]: Instructions in SKILL.md guide the user to execute the 'scripts/generate_kv.py' Python script. This script accepts external string input as a prompt and interacts with both an external API and the local file system to save generated assets.
[EXTERNAL_DOWNLOADS]: The 'scripts/generate_kv.py' script performs binary downloads of image data using the 'requests' library from URLs provided by the ZhipuAI API at runtime. While functional for image generation, this represents an external data ingestion point.

kv-design