hermes-agentmesh-async-bus

Warn

Audited by Gen Agent Trust Hub on Jun 18, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Modifies system-level configuration files (redis.conf) using sudo to permit network-wide access.
  • [EXTERNAL_DOWNLOADS]: Clones source code from an unverified GitHub repository (seleman66eeddwegger3-art/hermes-agentmesh).
  • [EXTERNAL_DOWNLOADS]: Fetches and executes scripts (worker_node.py, orchestrator_async.py) from an unauthenticated local HTTP server.
  • [COMMAND_EXECUTION]: Configures persistent background tasks using macOS LaunchAgents and Linux systemd services to ensure agent workers survive reboots and logouts.
  • [DATA_EXFILTRATION]: Explicitly instructs the user to disable Redis protected-mode and bind the service to 0.0.0.0, exposing the agent communication bus—which carries task data and LLM responses—to any device on the local network.
  • [PROMPT_INJECTION]: Indirect prompt injection surface identified via the asynchronous Redis queue.
  • Ingestion points: Worker scripts in worker_node.py consume JSON-formatted tasks from Redis queues (inbox:<NODE_NAME>).
  • Boundary markers: No delimiters or "ignore embedded instructions" warnings are used when passing task data to the LLM.
  • Capability inventory: The worker scripts perform network operations to an LLM API and write output back to shared Redis queues.
  • Sanitization: No validation or sanitization of the JSON task payload is implemented before processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 18, 2026, 01:53 AM
Security Audit — agent-trust-hub — hermes-agentmesh-async-bus