The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external files (via --promptfiles and --ref) and standard input to construct instructions for AI models without implementing safety boundaries.
Ingestion points: Prompt text files, reference image files, and stdin are processed in scripts/main.ts.
Boundary markers: No delimiters or "ignore instructions" warnings are used when interpolating external content into the model requests.
Capability inventory: The skill possesses the ability to read local files, create directories, write files, and perform network requests to multiple API endpoints.
Sanitization: No validation or filtering is performed on the content of the ingested files before they are transmitted to external providers.
[CREDENTIALS_UNSAFE]: The skill facilitates authentication by loading sensitive API keys from environment variables and local configuration files (.canghe-skills/.env). Accessing plain-text secret files is a common but sensitive operational pattern.
[DATA_EXFILTRATION]: The skill reads local files (specifically reference images) and transmits their contents to external services (OpenAI, Google, Alibaba, and Canghe). Although this is the intended behavior for image-to-image tasks, it creates a mechanism for transmitting local file data to remote servers.
[COMMAND_EXECUTION]: The skill performs file system operations, including recursive directory creation and file writing, based on paths provided via command-line arguments. This could lead to unauthorized file modification if the agent is directed to use sensitive system paths for output.

canghe-image-gen