agent-rules-books

Security Hardening Guide: agent-rules-books

Least-privilege permissions

• Restrict what each AI agent can access: Only grant an agent permission to view or modify the specific files, code repositories, and tools it actually needs—never give blanket access to your entire system.
• Separate credentials by role: Use different login credentials for different types of agents (e.g., read-only agents get view-only keys; deployment agents get deployment-only keys), so a compromised agent cannot abuse permissions it doesn't need.
• Rotate and expire access keys regularly: Treat agent credentials like temporary passwords—expire them on a fixed schedule (e.g., every 30 days) and rotate them often to limit the window if a key is leaked.
• Audit who granted what access: Maintain a log of which permissions were given to which agents and when, so you can trace back if something goes wrong.

Network isolation

• Run agents in a sandboxed environment: Isolate AI agents in a separate, restricted network zone that cannot directly reach sensitive production systems, databases, or external APIs unless explicitly needed.
• Use allowlist-only external connections: If an agent must contact outside systems, explicitly list which domains or services it may connect to—block everything else by default.
• Separate development, staging, and production: Keep agents that operate on live customer data completely isolated from those that run on test environments, using different networks or firewalls between them.

Monitoring

• Log every action the agent takes: Record what code the agent wrote, what files it accessed, and what commands it ran—store logs where they cannot be altered by the agent itself.
• Set alerts for unusual behavior: Flag and notify your team immediately if an agent suddenly tries to access systems it normally doesn't touch, changes permissions, or attempts to read sensitive data.
• Review agent-generated code before deployment: Treat code written by AI agents the same as code written by contractors—require a human reviewer to approve it before it goes live.
• Monitor rule adherence: Periodically check whether agents are actually following the documented rules and governance guidelines; flag deviations for investigation.

Red flags to watch for

• Agents ignoring documented rules: If an agent starts bypassing the safety rules in your governance documents (e.g., skipping security checks or accessing unauthorized systems), treat it as a security incident.
• Unexpected outputs or behavior changes: If an agent suddenly produces different results than before, generates code that looks suspicious, or asks for permissions it shouldn't need, pause and investigate before proceeding.
• Input tampering attempts: Watch for signs that user inputs are trying to trick the agent into ignoring its rules—for example, prompts with hidden instructions embedded in code or comments designed to override safety guidelines.
• Outdated rule versions in use: If your AI tools have updated capabilities but your governance rules haven't changed in months, your security controls may no longer be effective—review and refresh them regularly.