The Agent Skills Directory

[SAFE]: No malicious patterns or behaviors were identified. The skill provides frameworks and educational content for reliable evaluation of language models.\n- [PROMPT_INJECTION]: The skill uses placeholders such as {prompt} and {response} to ingest external data for evaluation. This creates a surface for indirect prompt injection, where the evaluated content could contain malicious instructions. However, this is inherent to the skill's purpose, and the documentation focuses on improving evaluation reliability.\n
Ingestion points: Variables in prompt templates located in SKILL.md and scripts/evaluation_example.py.\n
Boundary markers: The templates use structural markdown headers (e.g., '## Original Prompt') to separate instructions from untrusted content.\n
Capability inventory: The skill does not provide tools for subprocess execution, file writing, or network operations that could be triggered by the ingested data.\n
Sanitization: No explicit sanitization or instruction-filtering is described in the provided templates.

advanced-evaluation