The vulnerability affects OpenSSL 3.0.4, which was released last week. Furthermore, it only affects x64 systems with AVX-512 instruction set. Guido Vranken traces it back to his own May report on Montgomery modular exponentiation, whose fix for the dual 1024 RSAZ code went wrong (CVE-2022-2274). BoringSSL, LibreSSL and the 1.1.1 branch are not affected.
the reduction function is called with num set to the bit size, where it should be number of BN_ULONG elements (which are always 8 bytes large, because that is the size of an unsigned long on x64 systems, which is the only architecture which can have AVX512 support). So with the input sizes being 1024 bits, 8192 bytes are accessed (read from or written to) instead of 128 — Guido Vranken
He considers it trivially triggerable and sketches how the wrapping subtraction could bend function pointers towards code execution, or leak private data. Decipher adds that the fix was merged into OpenSSL master, but no release was scheduled yet. The fix shipped in OpenSSL 3.0.5.
By chance, Intel disabled AVX-512 on recent Alder Lake processors with a microcode update. Mainly affected is the Xeon processor family.