AgentInterdict
Integration guide · Hermes

Hermes Agent Security — enforcing agent safety in Hermes

Hermes agents combine persistent memory, tool access, and long-running autonomy with a messaging gateway that can act in the world. That's powerful — and it means untrusted content reaching the agent can try to steer real actions. This page is the Hermes-specific security guide.

Retrieval Permission

Hermes retrieves context and calls tools on your behalf. Retrieving a document or recalling a memory is not the same as authorising the action it suggests. That separation is what runtime enforcement provides.

Hermes' exposure

Hermes agents read untrusted content (emails, web pages, messages, documents), keep long-term memory, and call tools — including sending messages and making network calls. Without an enforcement layer, a single malicious document can attempt to override instructions and trigger an action.

The messaging gateway is the high-value target: a prompt injection that reaches a messaging-capable agent can try to post, send, or expose data to the wrong recipient.

Adding enforcement to Hermes

AgentInterdict integrates with Hermes and wraps the runtime boundary, intercepting memory writes and tool calls before they execute.

1
Memory writes scoredInstruction-shaped content from unverified origins is quarantined before persistence.
2
Tool calls revalidatedOrigin and authority are re-checked before a tool executes; externally-visible actions (send, post) from untrusted origins are blocked or gated.

Hermes security checklist

  1. Bound authority to origin — untrusted content never carries trusted-source privileges.
  2. Gate externally-visible actions (send, post, publish) behind action-time checks.
  3. Redact credentials before they reach memory.
  4. Revalidate at action time, not just at ingest.
  5. Fail closed when memory state is uncertain or tampered.

Related reading