The Agent Skills Directory

Accessing Local Authentication Tokens: The skill is configured to read a sensitive credential file from ~/.cache/huggingface/token using the cat command. Accessing local authentication files is a security consideration as it involves handling private tokens directly from the filesystem.
Transmission of Authentication Data: The retrieved token is transmitted to the Hugging Face API via curl headers for authentication. Although this targets the vendor's official domain (huggingface.co), the pattern of reading and sending credentials through shell command interpolation warrants review to ensure tokens are not inadvertently exposed in system logs or process monitors.
Processing External Third-Party Content: The skill ingests data from model cards and benchmark leaderboards (Step 3 and 4 in SKILL.md), which are ingestion points for third-party content. There are no boundary markers used to delimit this external data, and the skill lacks specific sanitization for the model metadata it processes. Given the capability inventory includes network operations (curl) and local tools (hf-cli, jq), this creates a potential surface for indirect prompt injection where malicious instructions in a model's description could influence agent behavior.

huggingface-best