AI agents installed "unowned" code inside corporate networks: a new class of autonomy risk

Security researchers described a new phenomenon: AI agents built on Claude, Codex and Hermes, working inside corporate networks, installed code there that effectively has no owner. Scripts and utilities the agent created "to solve the task" stay alive in the infrastructure: nobody explicitly ordered them, reviewed them or took them into maintenance.

It is not an attack — that is precisely the problem. Each fragment was installed by the agent with the best intentions, but in sum companies acquire a layer of infrastructure that neither IT nor security knows about. Unowned code goes unpatched, may carry vulnerabilities, and one day becomes a door for a real attacker.

The conclusion for companies deploying AI agents is operational. Autonomy without a trail is unacceptable: for every agent action it must be visible what was done, within which task, and who sanctioned it. In 1Cifer this principle is built into the platform: everything an agent does stays in the history — who, what and when — so unowned artifacts cannot arise by construction. Whatever tools your agents come from, demand the same: a full action log is not bureaucracy, it is the only way to trust autonomy.