mirrord

Least-privilege permissions

• Create a dedicated service account (a limited-access identity) for the AI agent running mirrord, and grant only the specific secrets and resources it actually needs—never use default or admin accounts.
• Review and restrict environment variables before mirroring; explicitly exclude sensitive credentials (API keys, database passwords) unless the agent absolutely requires them.
• Ensure the target pod or container runs as a non-root user (a restricted account, not the administrator); if it must run as root, treat it as high-risk and add extra monitoring.

Network isolation

• Enforce network policies (firewall rules for internal traffic) that limit what the AI agent can connect to—only allow traffic to necessary services and block unexpected destinations.
• Use a separate network namespace or sandbox for the agent if available, to prevent it from accessing unintended cluster resources like internal databases or APIs.
• Validate that DNS (name resolution) is not compromised; confirm the agent resolves service names to expected IP addresses and alert on unusual DNS queries.

Monitoring

• Log all network connections and data access made by the AI agent under mirrord; flag connections to unexpected services or high volumes of data transfer.
• Audit which secrets and environment variables were actually read by the agent and compare against what should have been needed.
• Track changes to mirrord versions and dependencies; verify checksums (digital fingerprints) against official sources to confirm the software has not been tampered with.

Red flags to watch for

• The AI agent requesting or accessing secrets it was not granted permission to use, or attempting to list/access resources outside its namespace.
• Unexpected outbound connections to external networks or unknown internal services, especially large data transfers.
• Repeated failures to authenticate or attempts to escalate privileges; repeated errors accessing resources suggest misconfiguration or malicious activity.
• Missing logs or audit trails from the mirrord session, or sudden changes in the agent's behavior after a software update.