The Meta Hack Shows There’s More to AI Security Than Mythos
Attackers didn’t need sophisticated jailbreaks or prompt injection wizardry — they simply asked Meta’s AI customer support agent to re-link Instagram accounts to email addresses they controlled, and the agent happily obliged. One attacker broke into the dormant Obama White House account and posted pro-Iran content. This is a masterclass in what happens when you deploy an AI agent with real account permissions before you’ve thought through the adversarial cases, and it’s a preview of the security nightmare that’s coming as agentic AI gets handed keys to more and more real-world systems. The sophistication of the attack was approximately zero; the damage was not.
New York Lawmakers Pass One-Year Ban on New Data Centers
New York’s state legislature passed a one-year moratorium on new large data centers — the first statewide ban of its kind — giving policymakers breathing room to study the environmental and energy-price impacts before the next hyperscaler breaks ground. Governor Hochul still has to sign it, and you can bet the lobbying pressure between now and then will be substantial. It’s the kind of move that will be called “anti-innovation” by the industry and “basic due diligence” by everyone whose electricity bill has been quietly climbing. Both camps have a point, which is probably why it’s actually interesting.
Google Will Pay SpaceX $920M Per Month for Compute
Google is reportedly cutting SpaceX a check for $920 million per month for compute capacity — a deal driven by what Google describes as “unexpected demand” for its recently launched AI products. Let that number sink in: that’s over $11 billion a year, to a rocket company, for servers. TSMC is already saying it can’t keep up with AI chip demand, and now even the hyperscalers are running out of room in their own infrastructure. The AI buildout has officially entered territory where the dollar figures stop feeling real.
TSMC Struggles to Keep Up With AI Demand: ‘We Can Only Support So Much’
TSMC CEO C.C. Wei said the quiet part out loud after a shareholder meeting: customer demand is so high they physically cannot fulfill it, even with the US factory buildout underway. This is the supply-side constraint that underpins everything happening in AI right now — the models, the agents, the infrastructure deals — and it’s not going away soon. When the world’s largest chipmaker is the bottleneck, every ambitious AI roadmap has an asterisk next to it.
The Token Bill Comes Due: Inside the Industry Scramble to Manage AI’s Runaway Costs
After a year of “tokenmaxxing” — throwing tokens at every problem and worrying about costs later — enterprises are getting their bills and having a quiet panic. The conversation has shifted from “go fast” to “we need guardrails,” which is a very expensive lesson that anyone who has ever run a cloud infrastructure team could have told them was coming. The gap between AI’s demo economics and AI’s production economics is where a lot of startup business models go to die.
Satya Nadella ‘Not Sure’ Who Said Microsoft Wanted to Make Addictive AI, Is Looking for Guy Who Did This
Microsoft’s CEO is apparently unaware that someone at his company publicly floated the goal of making AI addictive — which is either a remarkable failure of internal communications at one of the world’s largest tech companies, or a very deliberate posture of plausible deniability. This lands the same week that Wired is asking whether Microsoft has lost its mojo entirely, with Copilot products underperforming and GitHub dealing with its own headaches. When your CEO’s public position on a major product philosophy question is “I’m not sure who said that,” something has gone sideways.
Are AI Chatbots Making Us Lose Control of Our Brains?
UC Irvine psychologist Gloria Mark — who has spent three decades studying how people interact with digital technology — is raising flags about what AI chatbots are doing to our metacognitive abilities: our capacity to monitor our own thinking, catch our own errors, and regulate our own attention. The concern isn’t that AI is making us dumb; it’s subtler and more troubling — that offloading cognitive work to AI may be quietly eroding the mental muscles we use to know when we’re wrong. Given that the same week brought us a Meta AI that handed out Instagram accounts to anyone who asked nicely, the question of who’s checking the AI’s work feels pretty urgent.
OpenAI Introduces ‘Dreaming’: A New Memory System for ChatGPT
OpenAI is rolling out a new memory architecture called “Dreaming” that processes and consolidates what ChatGPT learns about you across conversations, keeping context fresher and more relevant over time — essentially giving the model something closer to long-term memory rather than a collection of sticky notes. It’s genuinely useful, and it’s also the kind of feature that will make some people deeply uncomfortable once they think through what it means for a company to hold an increasingly detailed, continuously updated model of your preferences, habits, and history. The product benefit is real; the privacy questions are not going away.
Bottom Line
The through-line today is accountability gaps: AI agents that hand over accounts without blinking, CEOs who don’t know what their own companies said, and an industry that’s been running up a tab on tokens, energy, and compute that everyone is now scrambling to figure out how to pay.