mission-control

Least-privilege permissions

• Limit credential scope: Create separate API keys for each runtime and external service with only the permissions needed for that specific tool—don't use master keys. If one key is stolen, only that service is exposed, not everything.
• Restrict who can dispatch tasks: Only authorized humans should be able to send work to your AI agents. Review who has "admin" or "task creator" access monthly and remove anyone who no longer needs it.
• Separate read and write access: Users who only need to review reports shouldn't have the ability to change settings, manage credentials, or launch new tasks.
• Use time-limited credentials: Set API keys and access tokens to expire automatically after 30–90 days. Force renewal regularly so old, leaked keys eventually stop working.

Network isolation

• Run on a private network: Deploy mission-control on an internal network or VPN, not on the public internet. Only your offices, authorized contractor locations, and specific CI/CD systems should connect to it.
• Firewall between runtimes: If your AI agents (OpenClaw, Claude Code, Codex) run on separate servers, use a firewall to block direct communication between them. This prevents one compromised agent from moving sideways to attack another.
• Encrypt credentials in transit: Use HTTPS/TLS (encrypted web connections) for all communication to and from the control plane, not plain HTTP.
• Restrict outbound access: Configure the server so it can only reach the specific external services it needs (e.g., OpenAI API, your databases). Block everything else to limit what an attacker can do if they break in.

Monitoring

• Log all task dispatches: Record every task sent to agents, including who created it, when, and what it does. Review logs weekly for suspicious patterns (e.g., tasks requesting data they shouldn't access).
• Alert on credential access: Set up automatic alerts if someone views, exports, or rotates API keys. Investigate any unexpected activity immediately.
• Track spend anomalies: Monitor your cost and usage reports. If spending suddenly spikes or patterns change dramatically, it may signal an attacker running unauthorized tasks.
• Audit permission changes: Log whenever access rules are modified (e.g., a new admin added). Review these changes monthly to catch unauthorized privilege escalation.

Red flags to watch for

• Outdated software: If mission-control hasn't received a security patch in 3+ months, or if the project has very few downloads, the tool may not be actively maintained. Plan to migrate or patch it yourself.
• Overly broad task requests: If an agent is suddenly asked to access files or services outside its normal scope, or if a task description is vague or unusual, pause and investigate before approving it.
• Unexplained credential rotations: If keys are rotated by someone you don't recognize, or rotations happen outside your normal schedule, assume a breach and rotate all credentials immediately.
• Lateral movement between runtimes: If you notice one AI agent accessing resources or data meant only for another agent, a configuration error or compromise may have broken the boundaries between them