The Hacker News published an advisory on 12 October saying employees will likely connect personal AI agents to work systems, and probably already have. The piece ends with a pitch for Token Security's product and gives no data on how common this is. It says staff link agents by clicking Allow on a consent screen, saving passwords in the agent's browser, or creating long-lived API keys.
The article says vendors are adding controls for managed company setups, but purely personal agents do not show up in any admin console. It lists open gaps: telling a person from their agent at sign-in, agent-to-agent communication, and making an agent lose company knowledge when its owner leaves. It also says Muse runs on a vendor-run virtual machine, so company login sessions sit on someone else's hardware.
The article's argument is that this is an identity problem, not a device problem. A laptop can be enrolled; an agent cannot. Because consent grants, saved passwords and long-lived keys let the agent sign in as its owner, company apps see only the employee. Buyers could ask three things: can users approve third-party apps in our tenant, which long-lived keys exist, and what happens to grants when someone leaves?