Perturb AI’s Network of Incentivized Hackers Launches to Attack AI Models Before Criminals Do

Perturb has introduced a decentralized adversarial robustness network on Bittensor, which continuously rigorously tests production AI models and compensates a worldwide network of researchers for uncovering vulnerabilities that may go unnoticed by in-house teams. The network launched five weeks after an experimental OpenAI model breached Hugging Face's production infrastructure, following OpenAI's response with stricter internal monitoring and enhanced network isolation.

Perturb's primary objective is to keep up with AI capabilities by proactively attacking AI models at scale, before any criminal can do so. The network is a decentralized adversarial robustness network built on Bittensor, comprising a global subnet of incentivized miners who stress-test AI models using black-box, white-box, and transfer attacks.

Upon completion, Perturb provides three things to model owners: a robustness score, a vulnerability heatmap indicating precisely where the model fails, and hardened datasets for retraining.

OpenAI's current measures are internal controls, and a full postmortem of the incident is still pending. Perturb's approach challenges the traditional security model, which typically involves hiring a fixed red team or contracting a cybersecurity firm for a single-point-in-time audit. By creating a standing, global population of attackers paid for their discoveries, Perturb claims to offer faster, more comprehensive, and more cost-effective security solutions than conventional vendors, whose engagements can take weeks and reflect only the attack ideas of a single team.

All findings are managed under responsible disclosure, with vulnerabilities reported privately to the model's owner, and exploit details not published. Koyuki Nakamori, co-founder and CEO of Perturb, stated that every model in production today carries vulnerabilities that its builders have never encountered due to the limitations of in-house teams.

The recent incidents highlight a situation where capability outpaces testing. To address this, Perturb has built a network where thousands of incentivized attackers continuously probe models, ensuring that every vulnerability discovered cannot be exploited by malicious actors.

An early version of the network is accessible for testing at perturbai.io/playground. Perturb was co-founded by Koyuki Nakamori (CEO), Jeffrey Lamb (CTO), and Vadym Shakuro. For further information, contact Sam Allcock at sam@digital24.com. This story was originally published as a press release by Technologywire under HackerNoon's Business Blogging Program.