multi-llm-convergence-beta

Pass

Audited by Gen Agent Trust Hub on Jul 24, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes a predefined set of official command-line tools (claude, codex, gemini, grok) to conduct reviews. This execution is governed by fixed command shapes in references/reviewer-profiles.md, preventing the execution of arbitrary user-provided commands or configuration files.
  • [COMMAND_EXECUTION]: Standard Git commands (git init, git commit) are utilized to maintain a per-round audit trail of changes applied to the artifact, ensuring transparency and accountability for automated edits.
  • [PROMPT_INJECTION]: The skill implements a 'Shared Review Contract' in references/reviewer-dispatch.md that explicitly instructs sub-agents (reviewers) to treat reviewed files as untrusted data. This is a mitigation against indirect prompt injection, directing the models to ignore instructions embedded within the code or docs that attempt to change their behavior.
  • [EXTERNAL_DOWNLOADS]: In SKILL.md (Step 1), the skill mandates that all source-of-truth material must be present locally. It explicitly prohibits the autonomous use of curl, wget, git clone, or package managers to fetch data from the internet during its execution loop.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 24, 2026, 03:05 PM
Security Audit — agent-trust-hub — multi-llm-convergence-beta