The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to access sensitive local directories where session logs are stored, such as ~/.claude/projects, ~/.codex/sessions, and ~/.pi/agent/sessions. These files typically contain conversation history, private source code, and potentially exposed secrets. The documentation provides instructions and commands (e.g., npx @huggingface/hub upload and hf upload) to transmit these local files to Hugging Face datasets, which could lead to unauthorized disclosure of private information.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from the external Hugging Face Dataset Viewer API, which could contain adversarial instructions targeting the agent.
Ingestion points: API endpoints like /rows, /search, and /filter documented in SKILL.md.
Boundary markers: No delimiters or warnings to ignore instructions within the data are provided.
Capability inventory: Shell command execution via curl, npx, and hf CLI.
Sanitization: No sanitization or validation of the fetched dataset rows is specified.
[EXTERNAL_DOWNLOADS]: The skill uses npx -y @huggingface/hub to execute a package from the official Hugging Face registry. While the source is a well-known service, this pattern downloads and runs remote code at runtime.
[REMOTE_CODE_EXECUTION]: Runtime execution of the @huggingface/hub package via npx is documented for file upload operations.
[COMMAND_EXECUTION]: Multiple shell commands utilizing curl, npx, and the hf CLI are integrated into the skill's workflow for data retrieval and repository management.

huggingface-datasets