huggingface-tool-builder

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the Hugging Face Hub (specifically model README.md files and metadata) which can potentially contain malicious instructions.\n- Ingestion points: The scripts references/hf_model_card_frontmatter.sh and references/hf_model_papers_auth.sh download and read model card content (README.md) from external repositories.\n- Boundary markers: No delimiters or "ignore embedded instructions" warnings are used when processing the fetched markdown content.\n- Capability inventory: The skill utilizes shell commands (curl, hf), Python parsing logic, and network operations to retrieve and process data.\n- Sanitization: The fetched content is not filtered or sanitized for potential injection patterns before being processed or presented to the agent context.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 03:52 AM
Security Audit — agent-trust-hub — huggingface-tool-builder