The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads local file content and sends it to the Anthropic API (via the anthropic library or claude CLI) for compression. This is the intended behavior of the tool. It includes a proactive is_sensitive_path heuristic check to prevent the accidental upload of credentials, environment files, and private keys to the third-party service.
[COMMAND_EXECUTION]: The skill executes the claude CLI tool using subprocess.run to handle LLM requests. The implementation uses a static argument list and passes data via standard input, which effectively prevents shell injection vulnerabilities.
[PROMPT_INJECTION]: The skill processes untrusted natural language files and interpolates their content into a system prompt for the LLM. This creates an indirect prompt injection surface where a malicious file could attempt to influence the LLM's output.
Ingestion points: Content is read from local files in scripts/compress.py.
Boundary markers: The file content is delimited by a テキスト: label within the prompt.
Capability inventory: The skill can read and overwrite local files and execute the claude CLI.
Sanitization: The tool implements a comprehensive blacklist of sensitive file names and paths (e.g., .env, .ssh/, secrets) to minimize the risk of sensitive data exposure.

genshijin-compress