Lilly Tech Systems
AI Agent Safe Coding Practices
Master the guardrails, permission models, and safety patterns that prevent AI coding agents like Claude Code, GitHub Copilot, Codex, Cursor, and Windsurf from accidentally destroying your infrastructure. Learn dry-run patterns, sandbox strategies, guardrail scripts, CI/CD safety, and incident response.
Your Learning Path
Follow these lessons in order, or jump to any topic that interests you.
1. Introduction
Why agent safety matters: the rise of AI coding agents, trust boundaries, real-world incidents, and why traditional security falls short.
2. Permission Models
How different AI agents handle permissions: Claude Code's allow/deny system, Copilot safeguards, least privilege, and sandboxed execution modes.
3. Dry-Run Patterns
Always preview before applying: terraform plan, kubectl diff, --dry-run flags, --what-if in Azure, and wrapper scripts that enforce plan-before-apply.
4. Sandbox Environments
Docker containers, dedicated cloud accounts, LocalStack, Azurite, GCP emulators, feature branches, ephemeral environments, and GitOps workflows.
5. Guardrail Scripts
Pre-execution hooks, blocklist patterns, Claude Code hooks, git hooks for IaC, shell wrappers, and OPA policy-as-code validation.
6. CI/CD Safety
GitOps workflows, GitHub Actions safety patterns, Terraform Cloud Sentinel policies, approval gates, branch protection, and automated plan review.
7. Incident Response
Containment, credential revocation, blast radius assessment, recovery from backups, post-incident review, runbooks, and communication templates.
8. Best Practices
Complete safety checklist, team policies, tool-specific guides for Claude Code/Copilot/Cursor, the Agent Safety Maturity Model, and FAQ.
What You'll Learn
By the end of this course, you'll be able to:
Configure Agent Permissions
Set up least-privilege permission models for AI coding agents so they can assist with development without having the power to destroy infrastructure.
Enforce Preview-Before-Apply
Build dry-run workflows and wrapper scripts that ensure every destructive operation is previewed, reviewed, and approved before execution.
Build Guardrail Scripts
Create pre-execution hooks, blocklist patterns, and policy-as-code rules that intercept and block dangerous agent commands automatically.
Respond to Agent Incidents
Follow proven incident response procedures to contain damage, recover resources, and build runbooks that prevent recurrence of agent-caused outages.