Cool to see a framework like AI-PTAF being proposed; definitely a step in the right direction. The main challenge, though, is that AI agents and the whole pentesting landscape are evolving at an insane pace, AI is practically shifting week by week.<p>So, for these certifications to actually hold weight and stay relevant, the benchmarks need to be truly living and adaptive. Think dynamic difficulty: if an agent solves scenario S1, then S1 itself (or the next scenario S2) should automatically adapt and become more challenging based on that successful performance. To achieve that level of real-time adaptation, the benchmarks themselves might need to be AI-generated, or hey, maybe just "vibe coded" by AI but fully adaptive in style, constantly evolving case-by-case to really push what these agents can do.
With the stochastic nature of LLM’s and the ever moving goal-posts, a framework not based off of knowledge that might already be in its memory is definitely needed