humanizer-ru

Fail

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONOBFUSCATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill instructions direct the agent to fetch and run a Python-based linter tool directly from a remote GitHub repository using the uvx command. This represents a remote code execution vector, as the fetched code is executed on the host system. While the command uses a specific commit SHA (e33188284ddd30b442ac86d91a00a79e3b3f3f2b) to pin the version, the security of this operation depends entirely on the integrity of the remote repository.
  • [OBFUSCATION]: A homoglyph was detected in the repository URL provided in the SKILL.md file. Specifically, characters in the URL (such as in the repository name or username) visually resemble standard Latin letters but are characters from a different script (e.g., Cyrillic 'е' instead of Latin 'e'). This is a known method for typosquatting, where an attacker impersonates a legitimate project or vendor to deliver malicious payloads.
  • [COMMAND_EXECUTION]: The skill explicitly instructs the agent to execute shell commands (uvx, humanizer-ru) to process files on the local file system. This capability, combined with the remote download, increases the risk of unauthorized system operations.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external Russian text files (текст.txt) to edit them for style and clarity. This ingestion of untrusted data creates a surface for indirect prompt injection attacks.
  • Ingestion points: The skill reads user-provided text files (e.g., текст.txt) for editing and auditing as described in the uvx command instructions in SKILL.md.
  • Boundary markers: There are no structural boundary markers or delimiters defined in the instructions to separate the input data from the agent's instructions during processing.
  • Capability inventory: The skill has access to shell command execution via uvx and file system read capabilities.
  • Sanitization: The instructions do not specify any validation, filtering, or sanitization of the input text before it is passed to the linter or processed by the language model.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Aug 28, 2026, 01:09 AM
Security Audit — agent-trust-hub — humanizer-ru