The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The file SKILL.md explicitly directs the execution of npx -y @nomadamas/k-skill@0 instruct korean-humanizer. This pattern downloads and immediately executes a package from the NPM registry. Since the @nomadamas account is not a recognized trusted vendor or associated with the skill author's verified infrastructure, this represents a high-risk arbitrary code execution vector.
[EXTERNAL_DOWNLOADS]: The skill requires the agent to fetch its 'full instructions' from a remote source (https://github.com/NomaDamas/k-skill/blob/main/korean-humanizer/instruction.md). Relying on unverified external repositories for core operational logic introduces a significant supply chain risk where the instructions could be modified maliciously without auditing.
[COMMAND_EXECUTION]: Multiple shell commands, including npx and git, are provided in SKILL.md for management and updates. These commands execute with local environment privileges and target untrusted third-party repositories.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and rewrite user-provided Korean text. The analysis of instruction.md reveals a lack of protective boundaries for this data. \n
Ingestion points: User-provided text for 'humanization' as defined in instruction.md. \n
Boundary markers: Absent. No specific delimiters or markdown blocks are used to isolate untrusted input. \n
Capability inventory: The agent has extensive instructions to modify, expand, or contract text, providing multiple avenues for an attacker to influence the final output through hidden instructions. \n
Sanitization: Absent. There are no instructions for the agent to filter or ignore potentially malicious prompts embedded in the text to be processed.

korean-humanizer