The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates an autonomous workflow where an LLM agent monitors training progress via alerts and metrics, creating a surface for indirect prompt injection. \n
Ingestion points: Untrusted data enters the agent's context through trackio list alerts --json and trackio get metric --json commands as defined in references/retrieving_metrics.md. \n
Boundary markers: No delimiters or explicit instructions to ignore embedded commands are specified when the agent processes retrieved data. \n
Capability inventory: The agent is authorized to launch background processes, adjust hyperparameters, and stop training runs based on the metrics, as described in SKILL.md. \n
Sanitization: No sanitization or validation of the alert text field is mentioned before the agent uses it for diagnostics. \n- [METADATA_POISONING]: The skill metadata contains a discrepancy between the provided author context ('sickn33') and the internal attribution to 'Hugging Face' in plugin.json. Additionally, the risk: critical tag in the SKILL.md frontmatter is misleading given the skill's primary functionality. \n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of the trackio Python package. This download targets the official repository of a trusted organization and is documented neutrally.

hugging-face-trackio