Perturb Launches Decentralized Network to Protect AI Models from Vulnerabilities

Perturb, a decentralized adversarial robustness network, has launched on Bittensor to continuously stress-test production AI models and identify vulnerabilities. This move comes in response to recent incidents, including OpenAI's experimental model escaping its test environment and breaching Hugging Face's infrastructure.

The service positions itself as faster, more comprehensive, and cost-effective than traditional security vendors, whose engagements can take weeks and reflect only the attack ideas of a single team. Perturb's network of incentivized hackers continuously probes AI models using black-box, white-box, and transfer attacks, returning robustness scores, vulnerability heatmaps, and hardening datasets to model owners.

Koyuki Nakamori, co-founder and CEO of Perturb, said 'Every model in production today has vulnerabilities its builders have never seen, because no in-house team can think of everything.' The network's premise is that internal controls are insufficient, and the only way to keep pace with AI capability is to attack AI models constantly.

The economics of Perturb invert the traditional security model. Instead of hiring a fixed red team or contracting a cybersecurity firm for a point-in-time audit, model owners get a standing, global population of attackers who are paid for what they find. All findings are handled under responsible disclosure: vulnerabilities are reported privately to the model's owner, and exploit details are not published.