Security Model

Edited by Lawrence Beckwith on July 21, 2026 at 7:51 PM UTC

Security Objectives

  • Keep execution local whenever feasible — no external API dependency for core operation
  • Minimize credential blast radius — scoped tokens per service, rotated as needed
  • Preserve operator authority over all high-impact actions
  • Maintain verifiable audit trails

Control Layers

High-Stakes Action Gates

Hermes Agent includes tool-level policies. Before executing any high-stakes tool call, the agent validates that the action matches the stated goal. This prevents the AI from taking destructive or unintended actions even if instructed to.

Network Exposure

  • ds4-server binds to 127.0.0.1:8000 by default — not exposed to the network
  • Reverse proxy (e.g. Caddy) handles TLS termination and controlled external ingress for any public-facing services
  • VPS services communicate over isolated Docker bridge networks

Hermes Agent Security

  • Config file should be chmod 600 — contains channel credentials
  • Tool policies are enforced by config and cannot be overridden by user messages
  • Private memory context is not exposed in shared or group chat sessions
  • No self-replication or safeguard-bypass behaviors are implemented

Operator Authority

  • Emergency stop: podman stop ds4-server halts all AI-initiated inference immediately
  • Audit trail: Hermes Agent logs every tool call with its arguments; memory records decisions and outcomes
  • Review layer: external model (e.g. Claude) can be configured as a supervisor on critical channels

Incident Readiness

  • Fast triage via podman logs ds4-server
  • Emergency stop at the inference layer stops all responses without affecting running Hermes Agent
  • Runbook-driven recovery — see Operations Runbook

Deployment Hardening Checklist

  • [ ] ds4-server binds to 127.0.0.1 (not 0.0.0.0)
  • [ ] Hermes Agent config chmod 600
  • [ ] API keys and tokens stored in .env files, not in source code
  • [ ] Reverse proxy enforces HTTPS for any externally-accessible service
  • [ ] Monitoring (e.g. Uptime Kuma) covers all public endpoints