The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves evaluating untrusted data (LLM outputs), which is a surface for indirect prompt injection where malicious content being judged could attempt to override the judge's instructions. The skill handles this appropriately by documenting the risk and advising users that prompts should not be treated as security boundaries.\n
Ingestion points: The skill ingests untrusted data through test_cases and the results of model.predict() in the evaluation suites and benchmark runner.\n
Boundary markers: Code snippets use basic string formatting and delimiters to separate instructions from data, which is standard for evaluation tasks.\n
Capability inventory: The skill possesses the capability to call external LLM APIs (e.g., OpenAI) to perform automated judging tasks.\n
Sanitization: While the code snippets do not implement specific sanitization, the documentation includes clear warnings about the limitations of prompt-based security boundaries during evaluation.\n- [EXTERNAL_DOWNLOADS]: The code snippets reference pre-trained machine learning models and libraries hosted on public repositories for legitimate evaluation purposes.\n
Evidence: Use of Hugging Face models (e.g., microsoft/deberta-xlarge-mnli) and the detoxify library for toxicity and groundedness checks.\n
Context: These are well-known, industry-standard resources from trusted organizations (e.g., Microsoft) used as intended within the scope of LLM evaluation.

llm-evaluation