ai-evals

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFENO_CODEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [NO_CODE]: The skill is comprised entirely of informational markdown files and does not contain any executable scripts, shell commands, or automation logic.
  • [SAFE]: The content includes descriptive analysis of AI security threats, such as prompt injection and artificial social engineering, specifically intended to educate users on identifying and defending against these attack vectors.
  • [EXTERNAL_DOWNLOADS]: The documentation references and recommends industry-standard AI evaluation tools and libraries from well-known technology organizations, including Arize Phoenix, Ragas, and Azure Responsible AI. These references are neutral and informative.
  • [INDIRECT_PROMPT_INJECTION]: The skill outlines methodologies for analyzing real-world user interactions and application traces. While processing untrusted data inherently creates an ingestion surface for indirect prompt injection, the skill proactively provides best-practice remediation guidance.
  • Ingestion points: Evaluation of user interaction traces and production data for error analysis (referenced in SKILL.md and artifacts.md).
  • Boundary markers: Recommends a structured "Four-Part Eval Formula" that defines specific context variables (e.g., {text}) to isolate untrusted data.
  • Capability inventory: No code-level capabilities such as subprocess execution, network operations, or file-system writes are present in the skill files.
  • Sanitization: Provides comprehensive frameworks for implementing "Guardrail Design Criteria" and "Constitutional AI" alignment processes.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 03:06 PM
Security Audit — agent-trust-hub — ai-evals