markitdown

Fail

Audited by Gen Agent Trust Hub on May 6, 2026

Risk Level: HIGHPROMPT_INJECTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill instructs the agent to modify its own persistent configuration files, which can affect the agent's future behavior across different sessions.
  • Evidence: The "Setup" section directs the agent to append a specific markdown block to ~/.claude/CLAUDE.md or ~/.gemini/GEMINI.md. This block instructs the agent to "auto-prefer" this skill whenever a file or URL needs to be read, effectively hijacking the agent's standard tool-selection logic.
  • [REMOTE_CODE_EXECUTION]: The skill documentation provides instructions for downloading and executing an installation script directly from the internet.
  • Evidence: In the "Error Handling" section, it suggests users run curl -LsSf https://astral.sh/uv/install.sh | sh to install the uv dependency.
  • [EXTERNAL_DOWNLOADS]: The skill dynamically fetches and executes code and containers from external sources.
  • Evidence: Uses uvx to run the markitdown package from Microsoft's official repository on PyPI. It also provides instructions for running a Docker container from ghcr.io/microsoft/markitdown:latest.
  • [COMMAND_EXECUTION]: The skill performs extensive shell command execution to handle file conversions and batch processing.
  • Evidence: Utilizes uvx, pip, and complex shell pipelines (e.g., find ... | xargs -P 4 ...) to manipulate the file system and process multiple documents.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, which is then processed by the agent, posing a risk that malicious instructions inside those files could be executed.
  • Ingestion points: Processes user-provided document paths and URLs (PDF, DOCX, XLSX, YouTube, RSS feeds) via the markitdown tool.
  • Boundary markers: The skill does not define specific delimiters or isolation markers when presenting the converted Markdown content to the agent.
  • Capability inventory: The agent has the ability to execute shell commands (uvx), write files to the system (-o flag), and perform network requests.
  • Sanitization: There is no evidence of content sanitization or instruction filtering for the text extracted from converted documents.
Recommendations
  • HIGH: Downloads and executes remote code from: https://astral.sh/uv/install.sh - DO NOT USE without thorough review
Audit Metadata
Risk Level
HIGH
Analyzed
May 6, 2026, 06:32 AM
Security Audit — agent-trust-hub — markitdown