openrig

Least-privilege permissions

• Give each agent only the specific access it needs. Don't share master passwords or admin credentials across your team of agents. Instead, assign each agent a limited role (e.g., "read-only reporter" vs. "data processor").
• Rotate credentials regularly. Change API keys and passwords every 30–90 days, and immediately revoke access for agents no longer in use.
• Separate sensitive data from shared context. Keep confidential information (customer data, financial records, secrets) out of the shared workspace where multiple agents can see it.

Network isolation

• Run agent teams in isolated environments when handling sensitive work. Use separate, dedicated deployments for high-risk tasks rather than mixing them with general-purpose agents.
• Limit external API calls. Restrict which agents can contact outside services (payment processors, third-party APIs) and log all outbound requests for review.
• Use a firewall or network boundary. Prevent agent infrastructure from directly accessing your internal databases or systems without explicit, monitored gateways.

Monitoring

• Log all agent conversations and decisions. Record what each agent does, what data it accesses, and what other agents it communicates with—keep logs for at least 90 days.
• Set up alerts for unusual activity. Flag unexpected credential use, large data transfers, or agents accessing resources they shouldn't, then investigate immediately.
• Audit agent outputs before deployment. Have a human review what agents produce before it reaches customers or affects business-critical systems.

Red flags to watch for

• Agents unexpectedly sharing sensitive information. If you notice credentials, personal data, or confidential instructions appearing in agent outputs or shared context, shut down the team and audit immediately.
• Requests from agents to bypass controls. If an agent asks to skip verification steps, ignore audit logs, or access restricted resources, treat it as a security incident.
• Unusual provider behavior or outages. Sudden rate limits, service degradation, or unexpected changes from your LLM provider (OpenAI, Anthropic, etc.) may signal compromise or third-party interference.