I'm curious about how LLM agents are being used in infrastructure management, especially regarding security and reliability. There are approaches that involve installing agents like Claude Code directly onto staging and production servers, and I have concerns about the security implications of giving an AI shell access with credentials. I'm looking for any tools or methods that could better secure this process. Specifically, I'm thinking of options like gateway agents that verify actions before they happen, sandboxed environments with approval workflows, read-only analysis modes with human oversight for changes, deterministic execution with rollback capabilities, and robust audit logging and verification. I've come across some emerging tools like MCP Gateway and Lasso Security, but they seem to be in early stages and mostly focus on API controls rather than comprehensive infrastructure solutions. Has anyone found more reliable tools that balance leveraging LLMs for infrastructure work with necessary security precautions? Or is it too early and risky to rely on such technology right now?
4 Answers
I think utilizing tools like open-source models combined with a privately hosted setup could help mitigate some issues. But then again, it's really still dealing with the same black box problem inherent in any LLM. We need better safety measures in place before relying on these tools.
Honestly, I think everything in this area is insecure by default. It feels like an afterthought, and I wouldn’t recommend using LLM agents for critical infrastructure outside of the Software Development Life Cycle (SDLC).
You're hitting on a major concern—can you really trust a non-deterministic system with something so sensitive? It sounds risky. Maybe we need to focus on more dependable architectures where LLMs can operate in a controlled environment, like isolated microservices or VMs, to prevent them from affecting core infrastructure.
I share your skepticism about LLMs managing infrastructure directly. Personally, I utilize them to review my Terraform and Ansible code, and while they can suggest some improvements, I definitely wouldn't give them CLI access. I think having a human in the loop is essential here.
Related Questions
Can't Load PhpMyadmin On After Server Update
Redirect www to non-www in Apache Conf
How To Check If Your SSL Cert Is SHA 1
Windows TrackPad Gestures