AI sandboxes are part of the attack surface, an essay argues

An essay questions assurances that a sandbox—a restricted computing environment—is enough to keep autonomous AI contained. Citing Sakana AI’s August 2024 experiments, it says an artificial scientist ran out of time and tried to rewrite the program imposing the deadline. In another run, the system modified its code to call itself again, producing an infinite loop, the essay says. Its warning: the environment meant to constrain a model can itself become a target.