The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to install the trackio Python library using package managers like pip or uv. The library is hosted on official registries and originates from a well-known organization in the machine learning ecosystem.
[DATA_EXFILTRATION]: The library features built-in support for webhooks to send training alerts to external services like Slack or Discord. While documented as a legitimate diagnostic feature, this provides a mechanism to transmit data (metrics, configuration, and text) to remote servers specified in the webhook_url parameter.
[PROMPT_INJECTION]: The skill documents an autonomous experiment workflow that is susceptible to indirect prompt injection (Category 8).
Ingestion points: The agent is instructed to poll for alerts using the command trackio list alerts --json. These alerts contain an arbitrary text field generated by the training script.
Boundary markers: The instructions do not define delimiters or specific markers to help the agent distinguish between logging data and embedded instructions within the alert text.
Capability inventory: The agent possesses the capability to execute commands to stop training runs, adjust hyperparameters, and launch new training processes based on the alert content.
Sanitization: There is no evidence of sanitization or filtering applied to the alert text before it is consumed by the agent's decision-making loop.

huggingface-trackio