The Agent Skills Directory

[DATA_EXFILTRATION]: The skill contains explicit instructions to access sensitive local directories containing agent session traces (~/.claude/projects, ~/.codex/sessions, ~/.pi/agent/sessions) and upload them to a remote service. This represents a significant risk of moving private interaction history to a third-party cloud environment.
[CREDENTIALS_UNSAFE]: The instructions acknowledge that these session traces can contain "secrets, or PII," yet still encourages the transfer of this data to Hugging Face repositories.
[COMMAND_EXECUTION]: The skill utilizes shell commands to manage data, including npx, hf upload, and hf repos create, which are used to move local files to remote servers.
[EXTERNAL_DOWNLOADS]: The skill invokes npx -y @huggingface/hub, which downloads and executes external code from the NPM registry at runtime.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by ingesting untrusted data from the Hugging Face Dataset Viewer API.
Ingestion points: Hugging Face Dataset Viewer API endpoints (/rows, /search, /filter, /parquet) in SKILL.md.
Boundary markers: Absent; there are no instructions for the agent to delimit or ignore instructions that may be embedded within the external dataset content.
Capability inventory: Subprocess execution via npx and hf-cli, sensitive local file system access (~/.claude/projects, etc.), and network operations via curl and the Hugging Face API.
Sanitization: Absent; the skill does not suggest any validation, filtering, or escaping of the content retrieved from the external API before it is processed by the agent.

huggingface-datasets