nemoclaw-email-policy
NemoClaw Email Policy
NemoClaw enforces network-level policies for AI agents. For email workflows, this means controlling which endpoints agents can reach — and which actions they cannot take — regardless of what the agent itself tries to do.
This is defense in depth: even if an email MCP server has a bug, or an agent hallucinates a send command, the network policy blocks the request before it reaches Microsoft's servers.
Why Policy Enforcement Matters
Email agents have access to sensitive data (inbox contents, contacts, calendars) and can take impactful actions (send email, create calendar events, modify inbox rules). In production, "the agent is well-behaved" is not a sufficient security model.
Failure modes without policy enforcement:
- Agent sends an email to an unintended recipient because of a hallucinated address
- A prompt injection in an email body tricks the agent into forwarding the message
- Agent creates an inbox rule that forwards all email to an external address
- A bug in the MCP server sends a draft that was not approved
NemoClaw prevents these by controlling the network layer — the agent process cannot reach endpoints that are not explicitly allowed.