We need to recognize that "agents" are not so much a technology as they are a fantasy.
Simply suturing together multiple LLMs, ass-to-mouth as if a sort of "human centipede," does not make for a functional technology.
This is clear.
More broadly, this isn't a "containment" issue, it's something inherent to the technology, viz. that there isn't any control to be had. What's happening here isn't a technology "going rogue," it's simply a nonfunctional fantasy technology failing to function.
If these AI firms can't even handle their own security controlling their own AI systems appropriately, why the bloody hell should ANYBODY trust their AI agents that they want everyone to give access to their work and personal lives? TOTAL INSANITY!