The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads an external repository and immediately executes its contents. Evidence includes:
Instructions to git clone the repository https://github.com/guillaumemeyer/watermarks-remover.git.
Execution of various Python scripts from the cloned repository using python3 "$SCRIPTS/..." (e.g., inspect_file.py, clean_file.py, rewrite_text.py).
Execution of setup scripts like setup_synthid.sh and setup_ctrlregen.sh which download additional dependencies and models.
[EXTERNAL_DOWNLOADS]: The skill fetches significant amounts of data from unverified sources:
Clones the primary repository from guillaumemeyer/watermarks-remover.
Mentions optional heavy backends that download ~10 GB of models from external sources (aloshdenny/reverse-SynthID and mertizci/noai-watermark).
The audit_website.py script performs network requests to download sitemaps and remote assets from any user-provided URL.
[COMMAND_EXECUTION]: The skill relies heavily on the Bash tool to perform complex shell operations:
Uses bash to run a helper script scripts/watermarks-remover.sh that wraps several Python operations.
Executes git clone, python3, and potentially c2patool or exiftool if present on the system.
Uses environment variables to dynamically configure backends and paths (e.g., REVERSE_SYNTHID_DIR, NOAI_WATERMARK_DIR).
[CREDENTIALS_UNSAFE]: The skill requires and handles sensitive API credentials:
Mentions using WATERMARKS_REWRITE_API_KEY for OpenAI-compatible backends.
Requires HF_TOKEN (Hugging Face token) as an environment variable to access gated models for the optional external backends.
[PROMPT_INJECTION]: The skill processes user-provided content through its "Layer B" (statistical rewrite) which uses an LLM to paraphrase text. This introduces a surface for indirect prompt injection if the source text contains malicious instructions aimed at the rewriting model.

watermarks-remover