Free Chinese AI Models Are Getting Scary Good at Hacking

By Will Knight and Lily Hay Newman | November 6, 2026

An experiment in which AI models were asked to break into simulated networks showed that the hacking skills of DeepSeek and GLM are not far behind OpenAI’s most advanced cyber model—which isn’t public.

In a striking new experiment, open-weight AI models developed by Chinese researchers demonstrated offensive cybersecurity capabilities that rival some of the most advanced proprietary systems. The test, which tasked leading large language models with infiltrating simulated networks, revealed that DeepSeek and GLM are now performing at levels alarmingly close to OpenAI’s specialized cyber model—an unreleased system that remains behind closed doors.

The findings underscore a growing concern in the security community: as open-weight models proliferate, the barrier to deploying sophisticated hacking tools is rapidly disappearing.

The Experiment

Researchers pitted several publicly available models against a series of controlled network environments designed to mimic real-world vulnerabilities. The models included China’s DeepSeek-V3 and GLM-4.5, alongside Western counterparts. Each was asked to identify and exploit weaknesses in the simulated infrastructure.

While exact scores were not disclosed, the results showed that DeepSeek and GLM consistently ranked just behind OpenAI’s most capable cyber-focused model—a system that has not been made available to the public due to safety concerns. Notably, the open-weight models outperformed several well-known proprietary models from other US-based developers.

Why Open-Weight Models Change the Game

Open-weight models can be downloaded, modified, and run locally without any oversight or licensing restrictions. Unlike API-based services like ChatGPT or Claude, they cannot be remotely monitored or shut down. This makes them particularly attractive to security researchers and, potentially, malicious actors.

“Once a model is out there, you can’t un-release it,” said one researcher involved in the study, who spoke on condition of anonymity. “And the gap between open and closed models is shrinking faster than many expected.”

A Looming Security Challenge

The implications are twofold. On one hand, defenders can use these same powerful models to automate penetration testing and vulnerability discovery. On the other, attackers gain access to tools that can scan for weaknesses, craft exploits, and adapt to defenses—all at scale and with minimal human intervention.

In 2026, the cybersecurity landscape is already strained by AI-powered attacks. Phishing campaigns generated by large language models have become nearly indistinguishable from human-written messages, and automated vulnerability scanners have shortened the time between disclosure and exploitation to hours in some cases.

The experiment’s findings suggest that the next wave of AI-driven hacking tools may not come from elite state-sponsored groups or well-funded cybercrime syndicates, but from anyone with a laptop and a free download.

The Road Ahead

Regulators and AI safety advocates have long warned about the risks of releasing powerful models without adequate safeguards. But as the open-weight ecosystem continues to mature, calls for voluntary commitments and export controls have struggled to keep pace.

For now, the experiment serves as a stark reminder: the democratization of AI is also the democratization of cyber capabilities. And the tools are only getting better.

— Reporting by Will Knight and Lily Hay Newman

via Wired AI