Control theory proves external AI safety enforcement impossible beyond a threshold

A new paper on arXiv (2605.12963) uses control theory to argue that once AI systems exceed a certain capability threshold, externally enforced safety strategies become structurally impossible. The authors prove two main results: first, under explicit premises including a reachability condition, any strategy relying on continued external enforcement cannot sustain safety once the system's effects outpace bounded control. Second, they outline structural requirements for any viable alternative, which must be intrinsic to the system rather than externally imposed. The paper does not name specific researchers or institutions, and no date is given beyond the arXiv submission.