coach-nemoclaw-hermes

Pass

Audited by Gen Agent Trust Hub on Jul 22, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes shell commands using the nemohermes tool to manage sandboxes and delegate instructions to the Hermes agent. These operations are performed within a controlled environment defined by the NEMOCLAW_SANDBOX_NAME variable.
  • [DATA_EXFILTRATION]: The skill manages the movement of data between host and sandbox environments via nemohermes sandbox upload/download. Potential risks are addressed through mandatory SHA-256 hash verification and file count checks to ensure integrity across boundaries.
  • [PROMPT_INJECTION]: The skill involves a delegation loop where Codex (primary agent) sends prompts to Hermes (sub-agent).
  • Ingestion points: Instructions are passed to the sub-agent via the $HERMES_PROMPT variable.
  • Boundary markers: Explicit isolation is defined between the Codex host, Hermes sandbox, and Blender MCP environments.
  • Capability inventory: Uses the nemohermes CLI for execution, file transfer, and status monitoring.
  • Sanitization: Employs SHA-256 hashing for data transfers and requires independent validation of artifacts and runtime receipts before accepting sub-agent claims.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 22, 2026, 03:08 PM
Security Audit — agent-trust-hub — coach-nemoclaw-hermes