multi-lens-quorum

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill orchestrates the spawning of multiple sub-agents to process queries. It references local file paths for these 'lenses' at /Users/engineer/workspace/backtest/.agents/skills/<LENS>/SKILL.md. This behavior is central to its stated purpose as a decision-orchestration tool and does not attempt to execute unauthorized system commands or escalate privileges.
  • [EXTERNAL_DOWNLOADS]: The skill enforces a strict citation rule requiring sub-agents to fetch external web content to ground their reasoning. It mandates the inclusion of a full URL, publication date, and verbatim quote for every external claim. This is a transparency and safety feature designed to mitigate hallucinations and ensure information is current.
  • [PROMPT_INJECTION]: The sub-agent prompt template includes explicit instructions to act as a 'context firewall,' directing agents to read only the assigned lens and ignore external data points not provided in the shared facts or fetched via the citation rule. This is a defensive pattern intended to prevent cross-contamination and indirect prompt injection from untrusted external sources.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 02:21 PM
Security Audit — agent-trust-hub — multi-lens-quorum