➥ AI security is entering the next phase, and @perturbaix is one project I'm keeping an eye on. Until now, adversarial testing has mainly been about finding attacks with internal teams or open-source tools. But Perturb is tackling something more interesting: continuous, incentivized hardening. ▸ Global miners compete to find real adversarial examples ▸ Verified attacks get used for model training, not just reports ▸ Robustness scores and datasets companies can act on ▸ Moving from subnet activity into an actual product This is the shift from one-time red teaming → ongoing external pressure that I think production AI needs. I'm currently watching how their product launch develops and whether teams outside the subnet actually use the output. #TAO #SN26 #Perturb