China's Open-Source AI Leaps Into the Global Top Four — and a Viral Story Gets It Wrong
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
A viral claim spread rapidly across social media Friday and Saturday alleging that Anthropic's AI system, called Mythos, had successfully hacked NSA classified networks. The story reached tens of thousands of shares before serious pushback emerged. According to the journalist who originally reported on Mythos and multiple industry leaders who responded over the weekend, the event was an authorized red-team test: Anthropic was invited by the government to attempt to penetrate classified systems specifically to identify vulnerabilities. The 'hack' framing transformed a structured and responsible security exercise into a panic narrative.
The speed of the story's spread illustrated how AI misinformation propagates differently from other types. Believers in imminent superintelligence treated it as confirmation of a capability threshold crossed; skeptics treated it as evidence of dangerous, uncontrolled AI development. Both camps shared it heavily — and both were wrong. The episode arrived the same week Elon Musk reiterated his prediction that AI will surpass all of humanity within four to five years, a timeline most AI researchers — including many working at frontier model companies — would privately dispute.
The competitive development demanding more sustained attention is China's. Z.AI, backed by Zhipu AI and closely connected to Chinese government research institutions, released GLM-5.2 as an open-weights model. It scored 51 on the Artificial Analysis Intelligence Index, placing it fourth globally and above every current Google model. Technology executives are reacting with what might charitably be described as genuine admiration and less charitably as alarm.
The open-weights dimension is decisive for understanding the competitive stakes. When a closed model such as GPT-5 or Claude 4 achieves a benchmark breakthrough, the capability remains within the company's infrastructure. When an open-weights model at near-equivalent performance is released, the capability becomes freely downloadable, deployable without API restrictions, and available for fine-tuning in ways the original developers never intended. GLM-5.2's benchmark ranking will influence enterprise procurement decisions in the near term.
A separate controversy highlighted the fault line running through AI training data. The artist SZA went public over the weekend after discovering that 238 of her songs had been used to train AI music systems without her consent. In contrast, Getty Images separately announced a licensing deal to display its images within ChatGPT — compensation in exchange for contribution rather than extraction without permission. Those two stories effectively map the industry's two diverging paths: systematic licensed compensation or continued operation in a legal gray zone that is generating an accumulating wave of litigation.