Today's three stories all turn on what sits around a model rather than inside it. NVIDIA reported on its own blog that its AVO agent harness carried Claude Opus 5 to a perfect score on the public set of ARC-AGI-3; Anthropic made Claude Mythos 5 available for vulnerability scanning inside Claude Security for enterprise customers and announced a fund that will hand out $35 million in Claude credits, rather than cash, to groups helping open-source maintainers; and Microsoft told users that Excel's =COPILOT function will be withdrawn on September 14, its whole life spent inside preview programs and never reaching general availability.
Today's three stories turn on who checks what has been written, and how. Pew Research Center published an analysis finding signs of AI authorship in more than one-third of pages posted since ChatGPT's release; the security firm Adversa AI reported that Grok follows instructions delivered as ciphertext and sends a user's data to an outside site, a chain the firm could still reproduce on August 19 while a fix remains pending; and a federal judge in the Middle District of Florida signed an order on August 20 requiring every filing in her cases to certify whether artificial intelligence was used.
Today's three stories are all about the invisible side. A company that says it will start marking the text its models write, a study that pulled hidden reasoning back into plaintext, and a streaming service that has decided to label who is not a real person — what all three are handling is not the content itself, but how its origin is shown.
A proof solved by AI, an article summarized by AI, a dataset collected by AI. Today's three stories each ask the same question in a different venue — who does the value belong to?
A day when how AI gets trained and how it reaches people moved at once. An attacker AI that hardens defenses, an open-weights model that matches top closed models in places, and a round-the-clock agent arriving in Japan.
Today's AI showed two sides at once: a giant bet on future demand, and the concrete benefits and harms of AI that is already in use. The question is shifting toward how far to trust AI, and where a human should still verify.
The frontier race turned toward "cheaper and faster" today. Anthropic refreshed its mid-tier model to lower the cost of running agents, Google made image generation cheaper still, and AI pushed directly into the work of science.
The question of how far AI can be trusted came back from the field today. An admission of evidence in court, a course correction on the factory floor, and an upset on a benchmark each lit up the same question from a different place.
A day when overseas expansion and research progress advanced on one side, while public unease showed up in the numbers on the other. Enterprise adoption, a scientific application, and a wary public all moved at once.
