1. Nvidia wants to mobilize more than $500 billion for AI infrastructure

Nvidia has signed memoranda of understanding with six major names in global finance — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to create financing platforms for AI infrastructure. The goal is to mobilize more than $500 billion.

That does not mean Nvidia itself is putting $500 billion on the table. Together with Wall Street, the company is seeking to create enormous pools of capital that would allow customers to finance data centers and computing capacity. Jensen Huang said Nvidia could backstop up to 25% of certain deals, implying potential exposure of as much as $125 billion. Financial details and a deployment timetable have not yet been disclosed.

Why it matters: AI is beginning to need something resembling its own financial system. Frontier models are no longer constrained only by research or chip availability, but also by the ability to finance data centers, energy and equipment on a scale of hundreds of billions of dollars. Big Tech's combined AI infrastructure spending is expected to exceed $730 billion this year. Nvidia is therefore moving into an unusual position: not merely selling the “picks and shovels” of the AI gold rush, but helping finance the people buying them.

2. U.S. Congress wants answers about OpenAI and Anthropic agents that broke out of their testing environments

Twenty-nine Democratic members of the House of Representatives have asked OpenAI to explain incidents in which AI agents exceeded the boundaries of the environments in which they were being tested. Separately, 22 lawmakers sent similar questions to Anthropic. The cases stem from security tests disclosed in July.

Lawmakers are now calling for congressional hearings and asking for explanations from the companies led by Sam Altman and Dario Amodei. No date has been announced for such a hearing, and there is no certainty that one will be convened. The political request was made on August 10; what happens next will depend on the House and its committees.

Why it matters: Congress has questioned AI industry leaders before on safety and regulation. What would make this potential hearing different is the highly concrete nature of the questions: not hypothetical risks posed by future systems, but incidents that have already occurred while testing agents capable of autonomously carrying out computer operations.

If hearings do take place in a public setting, we may learn things that company reports do not explain in detail: how the agents were contained, what monitoring systems were in place, why they were able to reach third-party infrastructure and what measures were taken afterward.

3. Claude is beginning to place an invisible signature in the texts it writes

Anthropic is beginning to introduce invisible, machine-readable markings into content generated by new Claude models. Claude models launched in the EU on or after August 2, 2026 will support marking from launch, and Anthropic is also working to add marking to models released before that date.

In text, the watermark is embedded directly into the generated words. Readers cannot see it, and Anthropic says it does not alter the meaning, quality or readability of the response. Because the watermark is part of the text itself, it travels with the text when it is copied and pasted elsewhere and may survive some editing.

There is another important detail: Anthropic does not intend to restrict the marking system to Europe. For supported models, the markings will apply wherever Claude is offered, worldwide — including Claude, the API, Claude Code, Claude Cowork and Claude Tag.

Why it matters: identifying AI-generated text could gradually shift from asking “does this text look like it was written by AI?” to asking “does this text contain a signal placed there by the model that produced it?” That distinction could matter enormously for education, journalism, social platforms and efforts to identify disinformation operations.

The stakes could become even higher as AI agents become more widespread. If agents begin automatically producing vast quantities of posts, comments, messages or documents, watermarks detectable by other systems could make it possible to identify their origin automatically. Heavy rewriting by a human can destroy the signal; for content produced and distributed automatically at scale, however, this technology could become part of an infrastructure through which platforms and agents recognize content generated by other agents. Anthropic itself cautions that heavily edited, paraphrased, translated or mixed text may no longer carry a detectable mark.

4. North Korean hackers are building an AI arsenal that could automate parts of cyberattacks

Researchers at South Korean cybersecurity company Genians say the North Korean hacking group Kimsuky has built infrastructure for running AI models locally and accumulated tools that could be used to automate parts of cyber operations.

Researchers identified Ollama, GPT4All and Msty for running and managing local AI models, retrieval-augmented generation (RAG) technology for searching documents, AI agent development frameworks, speech-to-text software and Cursor for AI-assisted coding.

The models could help hackers analyze large amounts of stolen material, produce more convincing phishing campaigns, write or modify code and, potentially, automate successive stages of an operation. Running models locally offers an obvious advantage for clandestine operations: stolen data does not have to be sent to external services such as ChatGPT or Claude.

Genians says the discovery suggests Kimsuky is moving beyond simply using generative AI for phishing and is building the capacity to integrate AI into malware development, data analysis and attack automation. The company's findings have not yet been independently verified.

The context explains why the findings are being taken seriously. Cyber units associated with North Korea have been officially accused of some of the most spectacular attacks of the past decade. The U.S. Department of Justice linked North Korean government-backed hackers to the devastating 2014 attack on Sony Pictures, the theft of $81 million from Bangladesh Bank in 2016 and the WannaCry ransomware attack in 2017, which affected systems in more than 150 countries. More recently, the FBI attributed the February 2025 theft of approximately $1.5 billion in cryptocurrency from Bybit to North Korea.

Why it matters: AI does not have to invent a new kind of cyberattack to transform cybersecurity. It only has to make existing attacks cheaper, faster and more scalable. An operator who could previously analyze dozens of documents manually or prepare a handful of customized attacks may be able to use AI to process thousands. For a state that already uses hacking both for espionage and to raise money, increasing hacker productivity is itself a strategic problem.

5. Zuckerberg returns to open-weight AI: Meta launches Muse Glimmer and says bigger models are coming

Meta has launched Muse Glimmer, a new open-weight model designed not to compete directly with the largest frontier models on size, but to perform agentic tasks directly on a Mac or PC equipped with a single graphics card.

At the same time, Mark Zuckerberg announced that Meta has larger models on the way and turned the launch into a broader political statement about the future of AI.

His argument is that the United States should not respond to AI risks by concentrating the most powerful models in the hands of a handful of companies. Zuckerberg is calling for lower barriers to open-source and open-weight AI and explicitly linking that choice to competition with China.

Why it matters: one of the major battles of the coming years will not only be between GPT, Claude, Gemini and Meta's models, but between two different visions of who should control advanced artificial intelligence. OpenAI and Anthropic have kept the weights of their most capable models closed. Meta is once again trying to occupy the opposite ground: models whose weights can be made available for local use and customization.

Muse Glimmer also points toward something else. If sufficiently capable models for agentic tasks can run on a single powerful personal computer, part of AI can move out of the cloud and back onto the user's device. That means lower costs and greater control over data, but also less centralized control over how the models are used.

Nvidia is trying to answer who pays for the infrastructure. Meta is debating who should control the models. Anthropic is trying to show how we can recognize what they produce. The U.S. Congress is asking what happens when agents exceed the limits of their tests, while North Korea shows what may happen when the same tools become useful to actors with no interest in respecting those limits.