eval-writer

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • Remote Dataset Acquisition: The skill provides templates for fetching evaluation datasets from external URLs. This is a standard pattern for integrating published benchmarks and is presented as a configurable setup step.
  • Controlled Execution via Sandboxes: For benchmarks that require running code (such as SWE-bench), the skill guides the agent to use sandbox providers like Modal or Daytona. This practice ensures that untrusted or generated code is executed in isolated, containerized environments.
  • Data Processing Surface: The generated evaluation suites are designed to ingest test cases from JSON fixtures or external APIs. While this introduces a surface for indirect prompt injection, it is the intended functionality for an automated testing framework and is mitigated by the suggested use of sandboxed execution for high-risk tasks.
  • Environment Variable Usage: The skill documentation references the use of standard environment variables for authentication with services like LangSmith and model providers, which aligns with best practices for secret management.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 08:06 PM
Security Audit — agent-trust-hub — eval-writer