The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation includes instructions for installing tools via shell pipes, such as curl | sh from astral.sh. It also references the trust_remote_code configuration option for HuggingFace datasets, which allows the execution of arbitrary Python code bundled with remote data.
[COMMAND_EXECUTION]: The skill documents capabilities for running shell commands, including cloud-based command execution (cli.cloud.modal_.run_cmd), local performance testing scripts (all_reduce_perf), and piped script execution (python scripts/cutcrossentropy_install.py | sh).
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes untrusted dataset content while maintaining powerful execution capabilities.
Ingestion points: Dataset loading modules described in references/api.md.
Boundary markers: No explicit boundary markers or safety instructions for the agent regarding embedded data content were found.
Capability inventory: Access to shell execution, file system operations, and network requests.
Sanitization: No sanitization mechanisms for external data content are documented.
[DATA_EXFILTRATION]: The skill facilitates data movement to external telemetry services (WandB, MLFlow, Comet) and various cloud storage providers (S3, GCS, Azure, OCI). While these are standard ML operations, they provide potential pathways for data exfiltration.
[EXTERNAL_DOWNLOADS]: The skill fetches content from external sources, including GitHub (cli.utils.fetch.fetch_from_github) and remote configuration URLs (cli.config.check_remote_config).

axolotl