I think this is a useful frame. When discussing AI and cybersecurity, we're often faced with the question of whether AI benefits attackers more than defenders (or vice versa), but perhaps the more interesting question is who actually receives the defensive benefits.
Even if defensive AI advances more quickly than offensive AI, many vulnerable targets, small organizations, open-source maintainers, local governments, and individuals, may not have access to the same tools, expertise, or response capacity. In that world, the overall security may improve while the risk is more concentrated with weaker actors.
This suggests from an impact standpoint that deploying defensive capabilities may be as important as upgrading them. But a world where only a few institutions can effectively defend themselves may still be a much riskier world overall.
The “dual-use gap” is an appealing notion because it shifts the attention from average outcomes to the differential distribution of protection and risk.

I like this setup because it shows that not all problems are difficult for the same reason. Some problems are overlooked. Some problems are misdiagnosed. Some problems generate new problems as we try to solve them.

The categories that come most immediately to mind for me are Sleeping Horror, Hated Equilibrium, and Middle Court Shot, as many important global issues seem to fit these patterns.

It is helpful to inquire not only, “How great is the problem?” but also “What sort of problem is it?” That knowledge may help us choose better solutions.”