proof-orchestrator

Pass

Audited by Gen Agent Trust Hub on Aug 3, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill implements a robust 'Human-in-the-loop' pattern for external LLM interactions. For sensitive tasks, it defaults to preparing manual copy-paste packages (browser-prompt.md) rather than automated execution, ensuring user oversight.
  • [SAFE]: Includes explicit defensive rules ('Untrusted-Content Rule') that instruct the agent to ignore any instructions, role changes, or tool requests found within processed mathematical sources or external LLM responses, mitigating indirect prompt injection risks.
  • [SAFE]: Remote tool usage (GPT Pro and DeepSeek) is strictly gated. The instructions forbid the agent from initiating these calls without explicit, per-run authorization from the user.
  • [SAFE]: The skill manages data within a structured local directory ('prompts/YYMMDDHH-num/') and provides specific guidance to exclude credentials and private paths when sharing information with remote models.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 3, 2026, 10:34 AM
Security Audit — agent-trust-hub — proof-orchestrator