hugging-face-evaluation

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). The extract-readme workflow loads and parses the target model card README content from Hugging Face (ModelCard.load(repo_id) → card.content → extract_tables_with_parser(readme_content)), so any outsider-authored README text (i.e., the repo’s README not authored by the operating user) can be ingested into the LLM context as free-form prose.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill runs evaluations that download Hugging Face model repositories at runtime (e.g., https://huggingface.co/{model_id}) and explicitly exposes/uses a --trust-remote-code option and backends that will load/execute code from those model repos, which can result in executing remote code fetched during runtime.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (medium risk: 0.60). The skill instructs running local evaluations (vLLM, lighteval, inspect-ai) and even explicitly recommends using the --trust-remote-code flag for models with custom code, which encourages executing untrusted remote code on the host (a clear risk of compromising machine state), but it does not request sudo, creating users, or modifying system/privileged files.