u2secured_team

There were obvious signs that things should look fishy, but I understand people cannot be vigilant all the time, and hence the trap. Once way to keep remining employees and staff is constant refresher courses. But this has been a problem haunting my workspace as well

Early in development phase, we realised vibecoding required more guard rails and also human checks. Had this problem where agent would try to use the ssh key , connection string.

Started additing permission deny to claude instructions helped, however, not 100% fool proof

I don't trust agentic systems with prod; rules always leave room for ambiguity and misinterpretation. There have been cases where agents have gone rogue, exposed data or even deleted production databases. Not worth the risk. This is where human software engineers are still required.