AI-Assisted DevSecOps Workflows
“Putting LLMs in the Deployment Pipeline”
Should an LLM read your infrastructure code? Should it suggest Terraform changes? Should it triage security alerts? This guide answers those questions. It is not a tutorial. It is a living document that records what works, what breaks, and what security controls are non-negotiable when AI assistants touch production systems.
What this is
A Hugo + Docsy documentation site that covers:
- Core architecture patterns for AI-assisted operations
- A comparison of LLM assistant frameworks (oh-my-opencode-slim, Aider, Claude Code, ShellGPT, AIChat)
- Security concerns specific to AI in DevSecOps: prompt injection, data exfiltration, over-privileged agent permissions
- Three paradigms: orchestrated multi-agent, single-agent pair programming, and CLI command generation
- Production-ready configuration templates for OpenCode Go
The zero-trust principles
AI assistants in DevSecOps must follow the same principles as any other automated system with production access:
Verify before execute. Every AI-generated command and configuration change requires human review. The assistant can propose, it cannot apply.
Least privilege per agent. Different tasks require different permissions. The agent that triages security alerts does not need access to modify Kubernetes deployments. The agent that writes Terraform does not need to read secrets. Scoped permissions per agent and per task.
Complete audit trail. Every interaction with an AI assistant is logged: the prompt, the context, the response, and the human decision that followed. If something goes wrong, you need to know exactly what the assistant saw and suggested.
Data sovereignty. Control what code and context leaves your environment. Self-hosted models via Ollama keep everything local. API-based providers require a clear policy on what data gets sent.
The three paradigms
Orchestrated multi-agent. Multiple specialized agents (Oracle for architecture review, Explorer for codebase search, Fixer for implementation) coordinated by a central orchestrator. Best for complex, multi-step security workflows that require different types of expertise. Used with oh-my-opencode-slim.
Single-agent pair programming. One AI assistant working alongside a human engineer in a focused session. Best for incident response, debugging, and development sprints where context switching is expensive. Used with Aider or Claude Code.
CLI command generation. The assistant generates shell commands on demand. Best for daily operations, troubleshooting, and rapid prototyping where speed matters more than depth. Used with ShellGPT or AIChat.
Production configurations
The repository includes ready-to-use configuration templates:
- A homelab GitOps config for K3s + OpenTofu + ArgoCD/Flux workflows
- An OpenCode Go preset with agent/model/skill/MCP routing
- A multi-provider DevSecOps config supporting OpenAI, Cerebras, and Fireworks
These are the configurations I use daily. They are opinionated and security-first by default.