ai agents are taking security proofs from handmade to mass production v12 and Zellic say it’s the first published proof covering a whole x86 program run, checked on the raw machine code the CPU runs, like inspecting the built bridge instead of the blueprint, and a second similar program took just 739 new lines. would be nice to get a checkable security guarantee about the compiled program itself, with fewer layers taken on trust. reusing the hardest parts means the next similar program takes much less work to verify, which could make rigorous security proofs practical at scale! it’s important to note that these are their own results, and they acknowledge their x86 model is still in development and hasn’t been audited by a human yet, but it’d be pretty cool if this scales https://t.co/CS5gv6j3uz