transformers-js

Pass

Audited by Gen Agent Trust Hub on Aug 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill documentation references the Hugging Face Hub (huggingface.co) and the jsDelivr CDN (cdn.jsdelivr.net) for model and library acquisition. Both are well-known, trusted resources for the JavaScript and AI communities.
  • [PROMPT_INJECTION]: The skill facilitates processing untrusted user data (text, images, audio) through machine learning models, creating a potential surface for indirect prompt injection.
  • Ingestion points: Untrusted data enters the agent context via pipeline() inputs in SKILL.md, references/EXAMPLES.md, and references/TEXT_GENERATION.md.
  • Boundary markers: No specific boundary markers or 'ignore instructions' warnings are included in the provided implementation examples.
  • Capability inventory: The skill has the capability to perform network operations (fetching models) and filesystem writes (caching), though these are restricted to library-defined paths.
  • Sanitization: No specific sanitization or filtering of external content is documented; the skill relies on the robustness of the specific ML models used.
  • [DATA_EXFILTRATION]: The skill describes best practices for managing API tokens (e.g., HF_TOKEN) via environment variables and accessing local cache directories like ~/.cache/huggingface/. These operations are transparently documented as part of the library's normal functionality.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 15, 2026, 03:43 AM
Security Audit — agent-trust-hub — transformers-js