AI agents have made destructive decisions on company codebases before, and (without proper guardrails) they will do it again.

Simply prompting agents not to take harmful actions isn’t enough. You need infrastructure-level controls and mitigations.

🧵👇