AI agents expose the trade-off between permissions, connections and realistic testing
The cases suggest that agent risk depends as much on the permissions and connections surrounding a model as on the model itself. They also show why fully isolated testing can fail to represent products that act across real services.