Performances
RD Popularity Index
RD Growth Index
| 1h | 24h | 7d | 30d | 60d | 90d |
|---|---|---|---|---|---|
| 0.54% | 3.27% | 11.06% | 52.91% | 60.92% | -- |
Comparison
About Perturb AI
Perturb AI (SN26) is a decentralized adversarial-robustness network on Bittensor, publicly launched in late August 2026. A global miner network continuously launches black-box, white-box and transfer attacks against production AI models; model owners receive a robustness score, a vulnerability heatmap showing where the model breaks, and hardening datasets for retraining, with all findings handled under responsible disclosure.
Mechanically, the scanning track gets about 80% of emissions and the training track about 20%; every 120 seconds the network picks new images from ImageNet-1k and distributes testing challenges, while validators assess attack quality, image fidelity, perturbation bounds and authenticity. It positions AI security testing as a continuously running, economically incentivized open market rather than one-off centralized red-teaming.
Milestone: Around October 12 the team released Whitepaper 3.0, outlining a decentralized AI security testing marketplace: the scanning track takes 80% of emissions and the training track 20%, with fresh tasks drawn from ImageNet-1k every 120 seconds; two commercial products are planned, an adversarial training dataset and model-robustness certificates, alongside paid robustness scans.
Market data: The subnet alpha fell about 40.6% over the last 30 days, with 909 holders, up 2.94% over 7 days.