Chinese open models near frontier hacking skills

An experiment that asked AI models to break into simulated networks found that free Chinese models from DeepSeek and GLM performed not far behind OpenAI's most advanced cyber model, which is not publicly available. The result is sharpening concerns that offensive security capability is spreading quickly through open-weight releases. Open-weight models can be downloaded and run without the API restrictions or monitoring that typically accompany hosted commercial systems. That openness has benefits for research and customization, but it also makes it harder to control how models are used. In the experiment, the Chinese models were asked to attack simulated networks, a controlled setting designed to test cyber capability. Their performance suggests that open models are approaching frontier-level skill in some offensive security tasks. Researchers caution that simulated environments are not the same as real-world intrusions. A model that can find weaknesses in a test network may still struggle with the complexity, unpredictability, and operational constraints of an actual target. Real systems involve messy configurations, live defenders, legal boundaries, and consequences that simulations cannot fully reproduce. The findings should therefore be read as a signal about capability diffusion, not as proof that open models can conduct sophisticated attacks end to end. Even with those caveats, the experiment adds urgency to debates over open-weight AI. Security teams want access to powerful models for defense and authorized testing, while governments and labs worry about misuse. The challenge is finding ways to preserve the benefits of open research without ignoring the risks. As cyber-capable models become more widely available, monitoring, responsible release practices, and clear norms for use will become increasingly important.