The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the official Kaggle CLI using the subprocess module with shell=False, which effectively mitigates shell injection vulnerabilities. It further implements an authorization layer in scripts/kaggle_runtime/commands.py to classify and restrict command groups, specifically blocking sensitive operations like printing access tokens.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: Robust redaction logic is implemented in scripts/kaggle_runtime/security.py to ensure that API keys, bearer tokens, and signed URLs from Kaggle's storage backends are stripped from all stdout, stderr, and audit logs before they are processed by the agent or stored.\n- [EXTERNAL_DOWNLOADS]: The skill manages dataset and metadata downloads from Kaggle's official API. It enforces strict path confinement and traversal prevention in scripts/kaggle_runtime/security.py, ensuring all downloads remain within an approved output root. Artifacts are verified using SHA-256 integrity hashes.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from Kaggle API responses. Ingestion points: Kaggle API results (datasets, kernels). Boundary markers: Absent. Capability inventory: Bounded file system write, Kaggle CLI execution. Sanitization: Automatic secret redaction. The attack surface is minimized by the wrapper's restricted command set and input validation.

kaggle-research