verbalized-sampling

Pass

Audited by Gen Agent Trust Hub on Jun 29, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements a multi-step pipeline where the output of the generation phase is directly fed into subsequent prompts for critique and judging. This creates a surface for indirect prompt injection.\n
  • Ingestion points: Data enters the context in references/critique-framework.md (Self-Critique Prompt) and references/judges.md (Judge Prompts) through placeholders like {idea_text} and [paste VS JSON output].\n
  • Boundary markers: The templates do not utilize distinct delimiters or instructions to the model to ignore any instructions embedded within the provided data.\n
  • Capability inventory: The skill is limited to LLM text generation and local Python script execution for formatting. No network exfiltration or sensitive file access tools are requested in the frontmatter or scripts.\n
  • Sanitization: The skill does not provide mechanisms to sanitize or escape the generated content before it is processed by the evaluation judges.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 29, 2026, 07:04 PM
Security Audit — agent-trust-hub — verbalized-sampling