Just to put a bit of context on those, OpenAI said they used 180,000 GPUs for training their GPT-5 model. As I understand it, each GPU has ~16,000 cores, and each of those cores does a couple billion operations per second. With that cluster, 2^65 operations takes a couple seconds, 2^90 takes many years, and 2^119 is off in “the sun has grown to surround the Earth” territory. The NSA is concerned with protecting their stuff against organizations with hundreds of billions of dollars, and a hundred of years of time to dedicate dozens or hundreds of people to breaking it. “Not good enough” in that context is a very different bar than what most people have to worry about.
You also need to provide them 2^43 unsalted responses, which is more than a casual few.
Just to put a bit of context on those, OpenAI said they used 180,000 GPUs for training their GPT-5 model. As I understand it, each GPU has ~16,000 cores, and each of those cores does a couple billion operations per second. With that cluster, 2^65 operations takes a couple seconds, 2^90 takes many years, and 2^119 is off in “the sun has grown to surround the Earth” territory. The NSA is concerned with protecting their stuff against organizations with hundreds of billions of dollars, and a hundred of years of time to dedicate dozens or hundreds of people to breaking it. “Not good enough” in that context is a very different bar than what most people have to worry about.
You also need to provide them 2^43 unsalted responses, which is more than a casual few.