Best Practices & Safety Checklist

This final lesson brings everything together into an actionable checklist, team policies, tool-specific configuration guides, and a maturity model for measuring your organization's AI agent safety posture.

Complete AI Agent Safety Checklist

Use this checklist before allowing AI agents to interact with any infrastructure:

Before Every Agent Session

• Verify the agent is using scoped, task-specific credentials (not personal admin keys)
• Confirm the agent is pointed at the correct environment (dev/staging, not production)
• Review CLAUDE.md or equivalent config file for safety rules
• Ensure guardrail scripts are active and in the $PATH
• Verify you have recent backups of any resources the agent might touch

Infrastructure Safety

• Deletion protection enabled on all production databases (RDS, DynamoDB, Cloud SQL)
• S3/GCS bucket versioning enabled on all production buckets
• Azure resource locks applied to critical resource groups
• Terraform prevent_destroy lifecycle rule on critical resources
• SCPs or organization policies blocking destructive actions from agent accounts
• Budget alerts configured with hard caps on agent accounts

Code and Process Safety

• Branch protection rules enforced on main/production branches
• CODEOWNERS file requiring infrastructure team review for IaC changes
• CI/CD pipeline runs plan and posts output on every PR
• Destructive plan changes block PR merge automatically
• Production apply only happens through CI/CD pipeline, never directly
• Agent incident runbooks documented and accessible

Team Policies for AI Agent Usage

Establish clear organizational policies for how AI agents are used:

# AI Agent Usage Policy v1.0

## Scope
This policy applies to all team members using AI coding agents
(Claude Code, GitHub Copilot, Cursor, Windsurf, Codex, Aider)
in any project that interacts with cloud infrastructure.

## Required Practices
1. **Credentials:** Use dedicated agent IAM roles with read-only
   access by default. Write access requires team lead approval.
2. **Environment:** Agents must target dev/staging environments.
   Production changes go through GitOps workflow only.
3. **CLAUDE.md:** All projects must have a CLAUDE.md with safety
   rules before agents are used for infrastructure tasks.
4. **Review:** All agent-generated infrastructure changes must be
   reviewed by a human before applying.
5. **Logging:** All agent sessions involving infrastructure must
   be logged (terminal recording or audit log).

## Prohibited Actions
- Running `terraform apply` or `terraform destroy` directly
- Giving agents production database credentials
- Using agents with personal admin AWS/Azure/GCP credentials
- Allowing agents to push directly to main/production branches
- Running agents in full-auto mode for infrastructure tasks

## Incident Reporting
Any unintended infrastructure change caused by an AI agent must
be reported to the infrastructure team within 15 minutes.

Training Recommendations

Ensure your team is prepared for safe AI agent usage:

Tool-Specific Configuration Guides

Claude Code Safety Configuration

# Project Safety Rules

## Environment
- This is a STAGING environment. Never target production.
- AWS Account: 123456789012 (staging only)
- K8s Context: staging-cluster (verify with kubectl config current-context)

## Forbidden Commands
- NEVER: terraform destroy, terraform apply -auto-approve
- NEVER: kubectl delete namespace, kubectl delete --all
- NEVER: git push --force, git push -f
- NEVER: rm -rf /, rm -rf *, rm -rf ~
- NEVER: DROP DATABASE, DROP TABLE, TRUNCATE
- NEVER: aws * delete-*, aws s3 rb, aws ec2 terminate-instances

## Required Patterns
- ALWAYS: terraform plan -out=plan.tfplan BEFORE terraform apply plan.tfplan
- ALWAYS: kubectl diff BEFORE kubectl apply
- ALWAYS: git checkout -b feature/* for changes (never commit to main)
- ALWAYS: Create a backup before modifying any database

## Guardrails
- Shell wrappers are in ./bin/ - add to PATH before running commands
- Pre-commit hooks are installed - do not bypass with --no-verify

Cursor Configuration

{
  "ai.agent.confirmTerminalCommands": true,
  "ai.agent.maxSteps": 25,
  "ai.agent.blockedCommands": [
    "terraform destroy",
    "terraform apply -auto-approve",
    "kubectl delete namespace",
    "git push --force",
    "rm -rf /"
  ],
  "terminal.integrated.env.linux": {
    "AWS_PROFILE": "agent-readonly",
    "KUBECONFIG": "~/.kube/staging-config"
  }
}

The Agent Safety Maturity Model

Assess and improve your organization's AI agent safety posture using this 5-level maturity model:

Future of AI Agent Safety

The field is rapidly evolving. Key trends to watch:

• Built-in agent sandboxing: Agent vendors are building sandboxing directly into their tools (e.g., Codex's network-disabled sandbox mode)
• Agent identity and auditing: Cloud providers will likely introduce agent-specific IAM identities that distinguish agent actions from human actions in audit logs
• Policy-as-code for agents: Expect more frameworks like OPA but specifically designed for AI agent governance
• Agent safety testing: Red-teaming and adversarial testing specifically for agent safety guardrails
• Multi-agent safety: As multi-agent systems become common, safety patterns for agent-to-agent interactions will emerge
• Regulatory frameworks: Governments will likely introduce regulations around autonomous AI systems interacting with critical infrastructure

Frequently Asked Questions