The Agent Skills Directory

[SAFE]: The skill is a documentation-only resource providing architectural and methodological guidance for designing trustworthy AI evaluation programs. It contains no executable scripts, binaries, or automated tasks that could perform unauthorized actions.
[EXTERNAL_DOWNLOADS]: The documentation references and provides integration snippets for several industry-standard evaluation frameworks and tools, including Hugging Face's lighteval, OpenAI's promptfoo, UK AISI's inspect-ai, and AWS Bedrock's evaluation services. These tools are sourced from well-known and trusted organizations.
[DATA_EXFILTRATION]: The skill emphasizes data privacy by instructing users to de-identify PII (Personally Identifiable Information) and sensitive data from production logs and support tickets before they are utilized in evaluation datasets.
[PROMPT_INJECTION]: The skill includes extensive guidance on identifying and mitigating security risks, with dedicated sections for safety red-teaming that cover jailbreak families, indirect prompt injection through tool outputs, and adversarial robustness testing.

ai-evals