Open weights: does the marginal risk argument still hold at frontier scale?

The standard defense is that open models only marginally advance what bad actors can already do. But as open models approach frontier capabilities, "marginal" starts doing a lot of work — the delta between closed and open shrinks while the downside of misuse stays fixed. I used to be firmly pro-openness and now I'm genuinely unsure. Where's the line where the calculus flips?