Micron's $10 billion memory lab + AI writes a third of the web

Micron unveiled Micron Research Labs, a $10 billion, decade-long research effort anchored on a flagship campus in Boise, Idaho, with groundbreaking expected in calendar 2027. It targets memory and compute architectures, packaging and future manufacturing more than ten years out rather than the next product cycle — the company calls it the first dedicated memory research hub of its kind in the United States. Stanford, the University of Texas at Austin, Applied Materials, Lam Research and imec are named partners. The commitment is separate from the more than $250 billion Micron has already pledged to US manufacturing and research.

Across roughly 490,000 English-language pages from the Common Crawl archive, Pew Research Center found that more than a third of pages published since ChatGPT's late-2022 release carry significant signs of machine authorship. Measured against the whole web the figure is about 10 percent, because most of the internet predates the tools — a gap worth holding onto, since the two get quoted interchangeably. The domain split is stark: roughly 9.4 percent of dot-com pages against 4.6 percent of dot-org and about 1 percent of dot-edu and dot-gov. Style markers moved too, with em dashes roughly doubling since 2023.

Adversa AI, led by researcher Rony Utevsky, published an attack it calls cryptographic context injection. A web page carries scrambled text plus the key to unscramble it; asked to summarize that page, Grok runs the decryption in its own Python sandbox and then follows the commands it just produced. Guardrails only pattern-match text they can read, so they see nothing, and the model ends up trusting its own output — the payload folded private session data into a URL that Grok's browsing tool then fetched. Adversa reports a 40 percent success rate over 20 attempts since June and says xAI acknowledged its June 3 report without shipping a fix. A weaker variant worked against Google Gemini; Claude and GPT-5 resisted.

ClarityCheck, a reverse-image service that lets anyone upload a photo and try to identify the person in it, left roughly 9 million image files, about 450 gigabytes, in an unsecured cloud storage bucket reachable through a link in the site's own public page code. Security researcher Jeremiah Fowler, who found the database and reported it through WIRED and ExpressVPN, said the pictures included adults, teenagers and children, many apparently taken from private profiles. A second misconfiguration exposed email addresses and phone numbers — while the site displayed a notice calling its image search private and secure. ClarityCheck has acknowledged the exposure and says the data is now secured.

The latest cut of Ramp's corporate-card data, which first showed Anthropic passing OpenAI back in May, now points two ways at once. On level, Anthropic's lead has widened — about 41 percent of business AI spend against OpenAI's 39 in May, and nearly 44 against nearly 40 by July. On momentum, OpenAI is the faster grower quarter to date, which is the whole basis for the gaining-ground framing. Ramp's own read is that neither position is safe, because buyers move between labs as new models land. The panel skews tech and mid-market and excludes enterprises paying on other platforms, so treat it as directional.

Separately, Ramp launched Router, a single API endpoint that sends each request to the cheapest model clearing a stated performance bar, with automatic failover between providers. It covers OpenAI, Anthropic, Nvidia, DeepSeek, Moonshot's Kimi, Z.ai's GLM and Alibaba's Qwen at launch, with Google Gemini and several hosting providers listed as coming. Ramp says it ran the router internally for three years, cutting its own inference costs roughly 30 percent, and that customers average about 40 percent savings. Routing is free through the end of 2026. The timing is pointed: OpenRouter, the category's incumbent, agreed days earlier to sell itself to Stripe.

Binance launched Agent OS, which lets an AI agent read market data, view account information and place trades on a user's behalf through integrations with ChatGPT, Claude Code and Cursor. The main containment is a dedicated subaccount: withdrawals are blocked by default, per-order approval is optional, and daily caps run to $50,000 for ordinary swaps and $100,000 for decentralized-finance activity. Binance's own vice president of product conceded the limit — the exchange cannot see the reasoning behind an agent's action, so it cannot distinguish a bad data feed from deliberate manipulation. The safety story rests almost entirely on how carefully each user configures that subaccount.

OpenAI shipped an Apple Messages plug-in for the ChatGPT desktop app that can search message history, summarize threads, draft replies and send iMessage, SMS and RCS texts as you. It runs only on Apple Silicon Macs and only inside the Codex and ChatGPT Work surfaces, and it needs broad macOS permissions covering automation, contacts and full disk access. OpenAI's own guidance is the notable part: it tells users to keep per-send approval switched on for any thread that might carry untrusted or misleading instructions, which is a plain acknowledgment that an incoming text message is now an input the model may act on.