The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/mux.sh is vulnerable to injection. It interpolates shell variables directly into a Python command string that is subsequently executed via uv run. If a file path provided to the script (such as the source subtitle file) contains malicious characters like single quotes or Python metacharacters, it could lead to arbitrary code execution.
[DYNAMIC_EXECUTION]: The skill uses uv run python -c "..." in scripts/mux.sh to assemble and execute Python code at runtime. The use of string concatenation and shell variable interpolation to build executable logic is a risky pattern that enables command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted subtitle text from user-provided files and passes it to LLM agents for translation and review in scripts/translate_workflow.js. This creates an indirect prompt injection surface where malicious instructions hidden in the subtitles could attempt to influence agent behavior during the style-guide generation or translation phases. The skill lacks robust sanitization or strict delimiters for this external data.
[EXTERNAL_DOWNLOADS]: The scripts/argos_translate.py script performs runtime downloads of language translation models from the Argos Translate project. Additionally, the skill uses uv run and nix-shell to fetch dependencies like the srt and argostranslate packages from public registries at runtime.

subtitle-translate