translate-book-parallel
Fail
Audited by Gen Agent Trust Hub on Sep 13, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill downloads and executes Python scripts (
convert.py,merge_and_build.py,calibre_html_publish.py) from an external, untrusted GitHub repository (github.com/deusyu/translate-book.git). This behavior allows arbitrary code from a third-party source to run on the user's system. - [PRIVILEGE_ESCALATION]: The skill's setup instructions direct the user to execute commands with administrative privileges (
sudo apt-get install calibre,sudo apt-get install pandoc) during the setup process. While these tools are common, requiring elevated privileges during skill installation is a security concern. - [EXTERNAL_DOWNLOADS]: The skill encourages downloading content from non-standard and untrusted sources, specifically using
npxandgit cloneto fetch the skill's core logic from thedeusyu/translate-bookrepository. - [COMMAND_EXECUTION]: The skill orchestrates multiple shell commands, including
ebook-convert,pandoc, and various Python scripts. This creates a significant execution surface that depends on the integrity of the downloaded scripts and external tools. - [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process external book files (PDF, DOCX, EPUB) and pass their content to parallel subagents for translation. This design is susceptible to indirect prompt injection, where malicious instructions hidden in a book could manipulate the subagents.
- Ingestion points: Book files are ingested and converted to Markdown chunks via
scripts/convert.pyas part of the translation pipeline. - Boundary markers: The skill documentation does not mention the use of specific delimiters or system instructions to prevent the LLM subagents from following commands embedded within the book text.
- Capability inventory: The skill possesses capabilities to execute subprocesses (
ebook-convert,pandoc), read/write to the local filesystem, and launch multiple parallel agent sessions. - Sanitization: There is no evidence of content sanitization or validation of the text chunks before they are sent to the translation agents.
Recommendations
- AI detected serious security threats
Audit Metadata