humanize-korean-ai-text
Warn
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONINGCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user to clone a repository from an unverified GitHub account (
epoko77-ai) as part of the installation process. - [REMOTE_CODE_EXECUTION]: The skill documentation suggests executing an installation script (
./scripts/install.sh) and installing plugins from an unverified source (epoko77-ai-plugins). - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text or files for rewriting, which presents a surface for instructions hidden in input data to influence agent behavior.
- Ingestion points: Text or file paths provided via the
/humanizecommand as described in SKILL.md. - Boundary markers: No explicit delimiters or instructions are defined to ignore embedded prompts in the ingested text.
- Capability inventory: The skill has the capability to write multiple files (JSON and Markdown) to a
_workspace/directory based on the processed content. - Sanitization: While the skill mentions safety rules (e.g., a 50% change limit), it lacks explicit sanitization or filtering for potentially malicious instructions embedded within the text to be humanized.
- [METADATA_POISONING]: The skill contains futuristic or potentially misleading metadata claiming to be from a "Daily 2026 Skills collection".
- [COMMAND_EXECUTION]: The instructions command the user to run shell operations including
git clone,claude, and script executions to enable the skill functionality.
Audit Metadata