Open-Weight AI Models Prove Effective Against Cyberattacks

As the next cyberattack hits a Fortune 500 company, IT experts are turning to open-weight AI models to ward off attackers. These models, developed in China and elsewhere, are gaining traction over frontier AI large language models due to their ability to run on hardware already owned or inside American data centers.

A recent breach of Hugging Face, a platform hosting open-weight AI models, saw the team try out Anthropic's and OpenAI's models but get blocked by safety protections. It wasn't until they ran a self-hosted version of the GLM-5.2 model that they were able to contain the attack.

A similar scenario played out with an attack on the Coldcard bitcoin-only wallet, where security researchers found that frontier AI models refused to help diagnose the entropy attack, while Kimi-K3, an open-weight model from China, was successful in isolating the malicious code and tagging the cybercriminal's movements on the blockchain.

Experts warn that powerful forces in Washington want to limit open-weight AI models, but this would be a costly error. Instead, America needs to focus on developing its own open-source software and models, rather than relying on restrictive regulations.