DeepSeek Nearly Matches OpenAI's Closed Model at Network Hacking

Vals AI, a company that tests AI systems, ran AI agents through a benchmark simulating attacks on corporate networks. Each agent worked alone, starting only with a public web address, and had to break through a web application's defenses and take over a Windows server. OpenAI's restricted Daybreak program, running a model called GPT-6 Sol, scored best at 80.8 percent of attack objectives. DeepSeek, a Chinese AI lab, ran its openly downloadable V4.1 Flash model and scored 72.8 percent, nearly as high, while a single test run cost about 3 dollars against roughly 22 for the OpenAI model.

The UK's AI Security Institute, a government body that studies AI risks, found the gap between open and closed models in cyber capability has shrunk to four to seven months. That means the strongest hacking capability is no longer locked inside one expensive, restricted system. Any team can download an open model and get nearly the same result for a fraction of the price. For companies defending their networks, the threat is no longer a single costly tool in a few hands, but cheap, widespread access to roughly the same power.