← sources

Do you consider giving your AI agent full access on your machine good OpSec?

source: reddit (reddit:devops)published 2026-11-18 21:36ingested 2026-11-19 02:36

retained text

Does your agent (Claude, Opencode, Codex, whatever) have access to your entire home directory? Can it use your ssh key, git credentials, docker credentials, cloud credentials etc.?

  
I have always considered these tools dangerous, because they could be following some random instruction from somewhere, so I have only been running them in sandboxed environments (vm's or containers) with their own readonly credentials.

However, this approach seems to be "holding me back" so to speak. I see colleagues who just YOLO their agent and are getting very good results. Agents that run commands on k8s clusters or remote servers through ssh, read logs and code, and can come up with solutions to problems very quickly.

So: I'm feeling conflicted. Keep the security paranoia or go with the flow and see what breaks?

What do you guys do?

document fetched 2026-11-19 00:25 · canonical https://reddit.com/r/devops/comments/1xzrlnt/do_you_consider_giving_your_ai_agent_full_access