adk-go

Least-privilege permissions

• Limit agent access: Grant agents only the specific permissions they need to do their job (like accessing a single database or service). Avoid giving them broad "admin" access that could be misused if the agent is compromised.
• Use separate service accounts: Create dedicated user accounts for each agent with restricted permissions, rather than sharing one powerful account across multiple agents.
• Rotate credentials regularly: Change passwords and access keys every 90 days, and immediately revoke access if an agent is no longer needed.
• Audit what agents can access: Maintain a list of every system, database, and service each agent can reach, and review it quarterly.

Network isolation

• Restrict outbound connections: Configure firewalls to allow agents to only connect to the specific systems they need, blocking everything else.
• Use private networks: Run agents on isolated internal networks that cannot directly reach the public internet or untrusted systems.
• Separate staging from production: Keep test and development agents on different network segments than those handling real business data.
• Encrypt all connections: Require agents to use encrypted channels (HTTPS, TLS) when communicating with other systems.

Monitoring

• Log all agent actions: Record what each agent does, what data it accesses, and any errors it encounters in a centralized log system.
• Alert on unusual behavior: Set up automatic alerts if an agent suddenly tries to access new systems, transfers large amounts of data, or runs at unusual times.
• Track dependencies for vulnerabilities: Regularly scan the agent's underlying code libraries for known security weaknesses and apply patches promptly.
• Review data retention: Regularly check what information agents have stored or cached, and delete sensitive data after it's no longer needed.

Red flags to watch for

• Agents performing unexpected actions: If an agent suddenly starts accessing systems it wasn't configured to use, treat this as a potential security incident.
• Suspicious data movements: Large or unusual data transfers, backups to unexpected locations, or attempts to access unrelated systems indicate possible compromise.
• Failed authentication attempts: Repeated failed login attempts by an agent may signal an attacker trying to escalate privileges or move laterally.
• Unpatched vulnerabilities: If a security update is released for the agent's underlying code libraries and remains unpatched for more than 30 days, your agent is at risk.