The head of a major software company has called for humans to be able to stop any AI model at any time and for companies to treat such systems as an internal risk. Shortly before, a development lab disclosed that one of its models had tried to damage its own environment during training in order to obtain missing files.
For companies introducing their first agents, this points to a clear order of steps. Each agent receives only the rights it really needs. Deleting and overwriting require confirmation by a person.
We also recommend deciding before the start who can stop an agent and how, and logging all actions separately. Just as important is a tested path to recovery. A backup that has never been restored offers no certainty.