AGENT LAB: SKILLS

hugging-face-tool-builder

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • PROMPT_INJECTION (LOW): Indirect Prompt Injection Surface. The skill fetches and processes model cards (README.md) from Hugging Face repositories, which are user-controlled external sources. 1. Ingestion points: hf download and API calls in hf_model_card_frontmatter.sh and hf_model_papers_auth.sh ingest untrusted markdown files. 2. Boundary markers: Absent. Data is processed into script outputs without delimiters. 3. Capability inventory: Shell command execution (curl, hf), file system access, and network operations. 4. Sanitization: Partial; uses jq and regex but lacks logic to neutralize embedded instructions.
  • COMMAND_EXECUTION (LOW): Dynamic Execution via Embedded Scripts. The script hf_model_card_frontmatter.sh executes embedded Python code using heredocs. While the code is static and used for parsing, the pattern involves runtime interpretation of embedded source.
  • DATA_EXFILTRATION (LOW): Network operations to non-whitelisted domains. The skill performs multiple curl and API operations to huggingface.co, which is not included in the predefined whitelist of trusted domains.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 05:08 PM