Free Chinese AI Models Are Getting Scary Good at Hacking (www.wired.com)

Recent research indicates that open-weight AI models are becoming nearly as proficient at hacking as their closed counterparts, signaling a major shift in cybersecurity dynamics. Nikil Rav from Vals highlights that these accessible models pose a significant threat as they match the capabilities of traditionally secured models like OpenAI's Daybreak. Specifically, in simulated hacking scenarios, DeepSeek, an open-source model, completed 73% of tasks while Daybreak achieved 81%. Notably, using these open models is far less expensive, costing between $2 to $3 compared to $17 to $46 for closed options, making them more accessible for potential attackers. This development raises urgent questions for cybersecurity strategies, as companies can no longer rely solely on closed models to maintain an edge over hackers. As cybercriminals increasingly harness AI tools, they are likely to amplify existing security challenges, necessitating immediate adaptations in defense mechanisms. Experts warn that while the rise of open models could democratize hacking capabilities, it also forces defenders to innovate and reinforce security measures. The unfolding situation presents a dual-edged sword—while it fuels a persistent arms race between attackers and defenders, it asserts that businesses must prepare for a landscape where advanced AI-driven hacking techniques are widely available.