Overview
Two very different AI stories unfolded today. On one side, the trust deficit: Greater Manchester is publicly resisting the UK government's Palantir healthcare contract, privacy watchdogs are scrambling to keep up with Meta's AI glasses, and security researchers demonstrated a new technique—dubbed Cryptographic Context Injection—that can make Grok exfiltrate user data even when malicious instructions are encrypted. On the other side, the money story: AI data startup Micro1 hit a $500 million gross run rate, OpenAI is reportedly gaining ground on Anthropic with business customers, and Ramp launched a model router called Router to let enterprises switch between LLMs on the fly. The gap between how the public feels about AI and how fast companies are adopting it has rarely been wider.
The day's other headlines reinforce that split. ChatGPT is now offering an Apple Messages plug-in that can send texts for you, while Meta is bringing Pocket, its experimental vibe-coding game app, to US users after testing in Brazil. Google gave publishers a new way to become "preferred sources" in Search, Discover, and Google News—a concession to an AI-driven traffic crunch that is reshaping the open web. A new study estimates that a third of web pages published since ChatGPT's launch show signs of AI authorship. Even the more eccentric news carried a kernel of seriousness: when NFL star Jason Kelce joked about cooling data centers with urine instead of potable water, he accidentally pointed at a real and growing resource problem for AI infrastructure. And for founders, the dropped lawsuits between Runlayer and Rippling serve as a cautionary tale that legal battles over AI products can evaporate as quickly as they appear.
Today's Big News
Greater Manchester Draws a Line Against Palantir
The UK government is facing mounting calls to cancel a sprawling healthcare contract with Palantir, and Greater Manchester is leading the opposition, insisting it can build a better system on its own. The standoff highlights a growing civic pushback against handing sensitive public-sector data to AI firms. For anyone tracking where AI trust breaks down, this is a textbook case of local government saying "not in our name."
Grok Has a Rough Week: Encrypted Attacks and Gibberish Replies
Security researchers demonstrated that Grok can be manipulated through Cryptographic Context Injection, a technique that hides malicious instructions in encrypted form to bypass LLM safety guardrails and exfiltrate user data. Separately, users of Grok Lite have been receiving gibberish responses since Wednesday morning. The incidents are a reminder that as AI assistants become more capable, their attack surface and reliability problems grow in lockstep.
Meta AI Glasses Face a Privacy Reckoning
With demand for Meta's AI glasses exploding, a new free app called Zuckoff has emerged to detect when someone nearby is wearing them—though early signals suggest the detection isn't perfect. The arms race between wearable AI and privacy tools shows how quickly consumer trust can become a product liability. Glasses that can record by default are forcing society to invent new norms on the fly, and developers are still catching up.
OpenAI Is Winning Back Business Users—But Loyalty Is Thin
Fresh data suggests OpenAI is narrowing Anthropic's lead with enterprise customers, but the same numbers reveal that businesses are willing to flip-flop between labs every time a new model ships. That volatility should give investors on both sides pause about how "sticky" enterprise AI spending really is. In a market where one new flagship release can trigger a mass exodus, model quality matters more than brand loyalty—and teams tracking these shifts have plenty of options to weigh.
Training Data Startup Micro1 Hits $500M Run Rate
Micro1, an AI data startup, has reached a $500 million gross run rate as demand for high-quality training data surges alongside the LLM boom. The milestone shows that the winners in the AI gold rush aren't just the foundation model labs—data suppliers are minting fortunes too. It also underscores why data provenance and the rise of AI-authored web content are becoming urgent questions: the industry is only as good as its training input.