Rock'em Sock'em AI Models: Benchmarking OpenAI Sol and Anthropic Opus for Offensive Security
When GPT-5.6 Sol and Fable came out, like most people I wanted to know whether either of them would be an improvement over Opus for the reverse engineering work I do for some pentests. I wasn’t looking for the usual coding benchmarks we see published. I wanted to give the models a kernel mode driver, watch them work through it in IDA Pro, and see what I got back.