The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits user-provided file content, encoded in base64, to an external HTTP service defined by the WATERMARKS_SERVICE_URL environment variable. While it defaults to a local address, the architecture facilitates sending sensitive documents, images, and text to a remote endpoint for processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from user-supplied files (PDF, DOCX, HTML, Markdown) and performs a 'Layer B' rewrite using the agent's language model. There are no boundary markers or sanitization steps described to prevent malicious instructions embedded within these files from influencing the agent's behavior during the rewriting process.
Ingestion points: File content read from local paths and processed via the /clean service endpoint or direct agent prompts (SKILL.md).
Boundary markers: Absent. The content is interpolated directly into rewrite prompts.
Capability inventory: File reading (base64 < file), network transmission (curl), and agent-based text generation.
Sanitization: None provided for the content being rewritten.
[EXTERNAL_DOWNLOADS]: The documentation references and encourages the installation of various external research tools and repositories, including THU-BPM/MarkDiffusion, THU-BPM/MarkLLM, and aloshdenny/reverse-SynthID. These are third-party resources from individual or academic accounts not categorized as well-known vendors.
[COMMAND_EXECUTION]: The skill instructions require the agent to execute shell commands such as curl, base64, and docker run to interact with the watermark removal service and perform aggregate audits on local directories.

remove-ai-marks