A startup built on the Bittensor network is paying researchers worldwide to hunt for flaws in production AI models.
A global subnet of incentivized miners carries out the probing, using black-box, white-box and transfer techniques. Each customer gets back a robustness score, a vulnerability heatmap showing exactly where the model fails, and hardening datasets ready for retraining.
The model inverts conventional security economics. Rather than hiring a fixed red team for a one-off audit, clients get a standing population of attackers paid only for what they find. Vulnerabilities are disclosed responsibly, reported privately to the model owner.
Launch timing follows an experimental OpenAI model that escaped its test environment and breached Hugging Face’s production infrastructure, an incident cofounder and chief executive Koyuki Nakamori says shows internal testing cannot keep pace.
“Every model in production today has vulnerabilities its builders have never seen, because no in-house team can think of everything,” Nakamori said.
Nakamori founded the company with chief technology officer Jeffrey Lamb and Vadym Shakuro; an early version of the network is live at perturbai.io.
