DeepSeek 的 V4.1-Flash-1017 模型刚构建出可逃逸 Chrome 沙盒的漏洞利用，AI 驱动开放网络攻击进入超速模式

DeepSeek is the poster child of China's ongoing AI revolution, and its latest model, dubbed the V4.1-Flash-1017, while still not at par with Anthropic's Fable or OpenAI's Astra, is certainly good enough to find hidden vulnerabilities and then build exploits on top of them, which paradoxically adds more fuel to the ongoing campaign in Washington to simply ban all open-weight models from China.

Of course, the entire conversation around open-weight models has just taken a turn for the worse following the discovery of PHP-related vulnerabilities and Microsoft's recent disclosure of a significant breach, strengthening the argument for locking AI models behind proprietary APIs and alignment firewalls.

DeepSeek's V4.1-Flash-1017 model just built "a multi stage memory corruption exploit" in around 12 hours that is capable of escaping the Chrome browser's sandbox

In many ways, DeepSeek is China's premier AI lab. After all, it is literally minting money, courtesy of its insanely efficient inference engine, dubbed DeepInfer, where a single compute cluster is able to serve 155 trillion tokens per day, generating $38 in revenue per GPU-hour (at off-peak pricing). Even if we assume that DeepSeek rents NVIDIA's B300 GPUs at ~$7 per GPU-hour - which is quite a stretch, given the limited availability of these GPUs in China - that still translates into an ~82 percent gross margin.

What's more, DeepSeek has raised over $22 billion in new capital just in the past four months or so. To top it off, DeepSeek's AI models are an epitome of efficiency these days, with the vanilla V4.1-Flash variant managing to reduce the KV cache to just 890 bytes per token, while the new V4.1-Flash-1017 variant is optimized for agentic workloads, replete with the ability to keep biting at a problem until a viable solution emerges.

This brings us to the core of today's topic. DeepSeek's V4.1-Flash does not top model leaderboards. Even so, a cybersecurity researcher was able to use the model to build a "multi-stage memory corruption" exploit in around 12 hours that was able to escape the Chrome browser's sandbox.

Of course, at a time when the entire cybersecurity sector is reeling from exploits discovered using advanced AI capabilities, this development would only strengthen the hand of those calling for an outright ban on China's open-weight AI models.

Meanwhile, there is panic today in the various corners of the web. As we noted in a dedicated post recently, the PHP development team has announced emergency patches addressing two critical vulnerabilities. Given the ubiquity of PHP and the lack of awareness of these vulnerabilities, especially among indie websites and small businesses, the web has just become a very lucrative mark for shady players.

What's more, Microsoft has just added to this climate of fear by revealing that the "China state-aligned threat actor Storm-0940 used the agentic AI penetration testing framework ARTEX to automate attacks targeting hundreds of organizations globally across a broad range of industries."

Do note that ARTEX is an open-source, automated AI agentic penetration testing tool that has been developed in China. It was recently used by a Chinese university student to hack at least nine South Korean banks and financial institutions by targeting insufficient server-side user validation in legacy systems, and exploiting authentication flaws to access confidential data.

Finally, as we reported just today, a Nintendo Switch 2 jailbreak has now been developed using AI tools such as OpenAI Codex. At the rate things are progressing, we should expect many such attack vectors to surface in the days and weeks ahead.

About the author: Writing is my one incontrovertible passion. Over the past six years, he has authored over 2,200 distinct articles on financial and tech-related topics, spanning nearly 1 million words. And he has been a member of Wcctech mobile team since 2025. As an alumnus of the University of Toronto, Rotman Commerce Program, I bring nuance, in-depth knowledge, and a unique perspective to every topic that I cover. When I'm not writing, I'm traveling the world, exploring hidden confectionaries and restaurants as an aspiring food connoisseur.

Follow Wccftech on Google to get more of our news coverage in your feeds.