Any AI pause will have defectors. How to ensure their incarceration actually prevents them from covertly contributing to AI research from behind bars?

To avert catastrophe, an international AI treaty is necessary. This treaty will, at minimum, need to ban the creation of artificial superintelligence, prohibit precursors beyond some threshold, and establish verification and enforcement mechanisms to allow countries to police each other. Much has been written on those matters, but less attention has been paid to what must be done with individual defectors - those who covertly seek to advance AI capabilities in defiance of the treaty ban - once detected and caught. Rogue AI developers pose unique challenges for the criminal justice system, because they have both the skills and the demonstrated motivation to advance the capabilities of systems that pose a catastrophic threat to humanity as a whole, and mere internet access - or even the ability to communicate with confederates outside prison - can allow them to pursue these goals. Because of this, such defectors will need to be handled in a special manner.

The linked essay focuses on non-state defectors, rather than defecting states, about which much more has been written. This might be a single individual running experiments on a massive cluster of illicit GPUs, a research team claiming to be running inference on existing AI models while secretly pulling off training runs for new models, or could come in a variety of different forms.