multi-llm-debate-grill-me

Pass

Audited by Gen Agent Trust Hub on Jul 14, 2026

Risk Level: SAFENO_CODE
Full Analysis
  • [SAFE]: The skill provides a set of structured instructions and templates for coordinating a debate between multiple Large Language Models. It does not contain any executable scripts, code, or configuration files that perform system-level operations.
  • [PROMPT_INJECTION]: The skill instructions are focused on creating a multi-round feedback loop between models. No patterns indicating an attempt to bypass agent safety filters, override system instructions, or extract system prompts were found.
  • [DATA_EXFILTRATION]: There is no evidence of hardcoded credentials, access to sensitive file paths, or unauthorized network operations.
  • [PROMPT_INJECTION]: The skill is designed to ingest and process external project briefs provided by users, which represents a potential surface for indirect prompt injection. 1. Ingestion points: The project brief section in PROMPT-ENTREE.md serves as the primary entry point for untrusted data. 2. Boundary markers: The instructions tell the model to not assume anything not written, but there are no formal delimiters to isolate or ignore instructions embedded within the user-provided brief. 3. Capability inventory: The skill does not include any tools or scripts that allow file system access, network communication, or shell command execution. 4. Sanitization: No input validation or sanitization is performed on the input brief. This surface is not exploitable due to the skill's complete lack of actionable functional capabilities.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 14, 2026, 11:31 AM
Security Audit — agent-trust-hub — multi-llm-debate-grill-me