multi-agent-eval-harness
Pass
Audited by Gen Agent Trust Hub on Jun 27, 2026
Risk Level: SAFENO_CODE
Full Analysis
- [SAFE]: The skill consists entirely of markdown documentation and conceptual pseudocode for a multi-agent evaluation framework. No executable code, scripts, or configuration files are included in the package. Analysis did not detect any malicious patterns such as credential harvesting, remote command execution, or data exfiltration.
- [PROMPT_INJECTION]: The skill outlines an indirect prompt injection surface inherent to evaluation frameworks where a judge LLM processes agent transcripts.
- Ingestion points: Agent transcripts and golden references in the
grade_transcriptlogic. - Boundary markers: Not specified in the documentation snippets.
- Capability inventory: The logic involves scoring content via an LLM; no system-level capabilities (subprocess, file-system access, or network operations) are included in the skill files.
- Sanitization: No validation or escaping of transcript content is described in the provided examples.
Audit Metadata