The Best Model Isn't the Threat. Cheap and Highly Capable Is.

For years, serious AI capability lived behind a wall of cost and controlled access. That wall is coming down now, and there are consequences for security teams.

I wrote a piece for the Arctic Wolf blog about GLM-5.2, a highly capable open model from the Chinese startup Z.ai. It runs close to the best proprietary models on long-horizon coding and agentic benchmarks and lands first among open models on Artificial Analysis's Intelligence Index. To be clear, it is not sitting at the frontier. It still trails the top closed systems on a number of benchmarks and some of what it can do may be distilled from those very models. None of that makes it less relevant. What makes it worth writing about is not that it is the best, but that something this capable is now cheap and open to anyone. The weights ship under an MIT license, and it costs are a fraction of what comparable systems charge. Anyone can run it for their own needs. That includes the people trying to break into their systems.

When capable AI gets this cheap, the cost of running a sophisticated attack drops with it. Reconnaissance and multi-step attack paths get faster to automate. You do not need frontier-level capability to do real damage with a model like this. AI is compressing the timeline of cyber risk, and manual triage was never built for that pace. I get into what that means for defenders, including where human judgment still sits at the center, in the full article on the Arctic Wolf blog.

Are you starting to see cheaper, open models show up in your own threat models yet? I'd love to hear how your teams are thinking about it.