Free Chinese AI Closes In on Elite Hacking Models

Chinese open-weight AI models are rapidly narrowing the gap with America’s most advanced systems in one alarming area: hacking into corporate networks.

In tests by Vals, an AI evaluation firm, as reported by WIRED, China’s DeepSeek completed about 73% of multi-step attack tasks on mock corporate networks. OpenAI’s restricted cyber model, Daybreak, completed about 81%.

The tasks were demanding, including breaching one network and then moving on into a second. On some individual tasks, the two models tied.

The bigger shock is cost. A full attack run cost just $2 to $3 with DeepSeek, against $17 to $46 with OpenAI’s top tier, roughly ten times cheaper.

Another Chinese model raises further concern. In September, Anthropic reported that GLM-5.3, built by Beijing lab Z.ai (formerly Zhipu AI), could carry out the steps of an attack on its own, from scanning for weaknesses to writing working exploit code, with little human help.

Anthropic also found that simple jailbreaks bypassed GLM-5.3’s safety filters 64% to 100% of the time in simulated tests.

US frontier models are increasingly locked behind identity checks, but these open models can be downloaded freely. Versions stripped of their safety training circulate with no logging or abuse detection.

For frontier developers, the trend points to diminishing returns. Spending billions to stay ahead buys a shrinking lead, as cheaper open models reach near-parity within months.

Caution is still needed. Benchmark scores do not prove real-world success, and performing well in a simulation does not guarantee a model can breach live systems. Most attackers still use older tools, and defenders can also run open models on their own infrastructure to find weaknesses first.

Still, the message for businesses is clear: advanced hacking capability is becoming cheap and widely available, and every security strategy must now assume attackers have it.