OpenAI dumped hundreds and hundreds of solutions with Lean proofs that hadn't been seriously checked by any human and AFAIK, none of them have turned out to have exploited bugs in the prover. Some of them didn't prove what their paper proof claimed they did, but the base rate of LLMs exploiting such bugs (which are definitely available to exploit -- even for a human it historically hasn't been SUPER hard to find them, and the Lean kernel's metatheory is suspect in a number of ways) is empirically very low, even when the incentives for a significant result are very high. The more likely pathologies are, e.g., misstating the problem (the Riemann hypothesis is using the standard Mathlib definition), adding extra unjustified assumptions (unlikely in this case -- I haven't inspected the Lean proof yet but others have and reported that there aren't any), or not matching the paper proof and intuition (I'd strongly suspect the latter is true in this case).