📝Perturb has been running continuous adversarial scanning and training for months. It might be one of the more practical AI security experiments on Bittensor so far. @perturbaix went live earlier this year and has already moved past pure attack discovery. According to recent updates: - Adversarial scanning and training tracks both live - Miners can now fine-tune models using the attacks they generate - Product phase officially starting after ~5 months of subnet work - Robustness scores, datasets, and certificates as the core outputs – Alpha buybacks and owner lockups already happening – Targeting a commercial launch in the coming weeks What’s driving the approach? Perturb turns finding model failures into an incentivized competition, then uses those verified adversarial examples for actual hardening instead of just publishing a report and stopping. The idea is continuous external pressure rather than one-off audits. I can see why teams working on production AI might pay attention. However, the setup still deserves a closer look. Subnet activity and attack success rates don’t automatically mean enterprises will buy the product, and scaling beyond current vision models will take real work. Also, emissions and early traction don’t guarantee sustainable revenue once the commercial layer is live. I’m keeping an eye on how the product launch actually converts. If Perturb can turn continuous adversarial data into something companies reliably pay for and use, this could become a meaningful experiment in decentralized AI security. #Bittensor $TAO #Perturb