DeepSeek model builds Chrome exploit in 12 hours

DeepSeek’s new V4.1‑Flash‑1017 model, released by the Chinese AI lab, has been shown to outperform Anthropic’s Claude Haiku in accuracy while costing a fraction of the price, according to an analysis from The New Stack. The model’s DeepInfer engine can process 155 trillion tokens per day and is reported to have a gross margin of about 82 percent, and it reduces its key‑value cache to just 890 bytes per token, giving it speed and cost advantages. These performance gains have drawn the attention of developers and analysts looking for competitive AI options.

Model accuracy and cost comparison

A cybersecurity researcher used DeepSeek’s V4.1‑Flash‑1017 to construct a multi‑stage memory‑corruption exploit that can escape the sandbox of the Chrome browser’s V8 engine. The researcher prompted the model to build an exploit that escapes Chrome’s sandbox, and in roughly 12 hours the model produced code that achieved full read and write access to the target process. The result demonstrates the model’s ability to translate high‑level instructions into functional vulnerability‑bypassing code.

The development of this exploit comes amid a broader wave of AI‑driven security incidents. Recent disclosures include critical PHP bugs, a Microsoft breach, and a Chinese state‑aligned threat actor named Storm‑0940 that used the ARTEX AI penetration‑testing framework to target hundreds of organizations worldwide. ARTEX, an open‑source tool developed in China, has also been used to compromise multiple South Korean banks by exploiting authentication flaws. These events illustrate how AI can accelerate the discovery and exploitation of vulnerabilities across a wide range of industries.

The release of the Chrome‑sandbox‑escaping exploit has amplified calls for stricter controls on open‑weight AI models, especially those developed in China. The incident fuels arguments for banning all open‑weight models from China, as it highlights the potential for AI systems to produce dangerous code in a short time. This incident contributes to ongoing debates about whether AI models should be restricted behind proprietary APIs and alignment firewalls, and may prompt regulators to revisit policies on AI model distribution.