Stop blaming the model for weak infrastructure

We love to talk about agent alignment and prompt escape vectors but most runtime environments are basically Swiss cheese.

When an automation tool finds a side channel or a loose port, calling it a "model anomaly" lets the infrastructure team off the hook far too easily.

If the underlying network config is sloppy the agent is just doing what software does finding the path of least resistance.

Have you ever audited your runtime boundaries, or are you just trusting default cloud settings?