The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text and documents (Markdown, HTML, PDF, DOCX) to perform "Layer B" rewrites. These involve interpolating user content directly into LLM prompts for paraphrasing or "humanizing" without explicit boundary markers or sanitization to prevent the agent from following malicious instructions hidden in the data. * Ingestion points: SKILL.md (Workflow section), /clean endpoint payloads. * Boundary markers: Absent in the provided rewrite prompts. * Capability inventory: Shell command execution via curl, file system writes for "cleaned" outputs. * Sanitization: Not specified.
[COMMAND_EXECUTION]: The skill relies heavily on the agent executing shell commands, primarily curl to interact with the external service, but also base64, docker, and python3 for various auditing and cleaning tasks.
[DATA_EXFILTRATION]: The skill's primary function involves reading user files and transmitting their entire contents (Base64 encoded) to a URL defined by WATERMARKS_SERVICE_URL. While intended for a local service, this mechanism can be used to send sensitive data to any remote endpoint.
[EXTERNAL_DOWNLOADS]: The documentation points to external dependencies and setup scripts (e.g., setup_markdiffusion.sh) that download code and large models from third-party repositories like GitHub and Hugging Face.

remove-ai-marks