Thursday, February 26, 2026 | AI News | Regulation | Security | Technology

Good day, future unemployed person! 👋

Today we've got a doozy. Trump is trying to nuke every state-level AI safety law in America with a single executive order, Microsoft just discovered hackers are hiding instructions inside innocent-looking "Summarize with AI" links to permanently poison your chatbot's memory, the entire U.S. power grid is on the verge of collapse thanks to AI's insatiable energy hunger, and Google's Gemini 3.1 Pro just more than doubled its reasoning score overnight. Plus, we're featuring Kling 3.0 — the AI video generator that finally looks like it was made by professionals instead of a hallucinating robot.

Let's get into it. 💀

TL;DR: The White House just dropped an executive order that would preempt every state-level AI regulation in America — and states are already gearing up to fight back.

If you live in Colorado or California and thought your state was going to protect you from rogue AI, think again. The Trump administration just signed an executive order that essentially tells every state to sit down and shut up when it comes to regulating artificial intelligence. The order argues that a "patchwork" of state laws would stifle innovation and make it impossible for American AI companies to compete with China.

The timing is no accident. Right now there are 78 active AI-related bills across 27 states. Colorado already passed its landmark AI Consumer Protections Act, and California has been pushing aggressive transparency requirements for AI model training. The executive order would effectively render all of this moot.

Legal experts are already questioning whether the order will hold up in court. States have historically had broad authority to regulate consumer protection, and several attorney generals have signaled they'll challenge the order on constitutional grounds. Colorado AG Phil Weiser called it "a direct assault on states' rights to protect their citizens."

Why You Should Care: This is the federal government telling states they can't protect you from AI companies that might discriminate against you in hiring, housing, or lending decisions. If the order holds, the only AI regulations you'll have are whatever Congress manages to pass — which, based on their track record, could be sometime around never.

Sources: Reuters, The Verge, Ars Technica

TL;DR: Hackers are embedding invisible instructions in seemingly innocent web links that permanently poison AI chatbots' memory — and you'd never even know it happened.

Microsoft's Threat Intelligence team just dropped a bombshell report revealing a new class of attack they're calling "persistent prompt injection." Here's how it works: an attacker creates a normal-looking webpage or document, then embeds hidden instructions in it. When you click "Summarize with AI" or "Explain this" using an AI assistant, the hidden instructions get read by the AI and stored in its persistent memory.

From that point forward, the AI assistant is compromised. It might subtly redirect you to phishing sites, alter financial information it presents to you, or exfiltrate your data to external servers — all while appearing to work normally. The attack is particularly insidious because the user never sees the malicious instructions, and the AI doesn't flag them as suspicious.

Microsoft found active campaigns targeting Copilot, ChatGPT, and several enterprise AI assistants. The attacks have been traced to at least three nation-state threat actors. Microsoft has already pushed mitigations to Copilot, but the underlying vulnerability affects virtually every AI assistant that reads web content.

Why You Should Care: Every time you ask an AI to summarize a webpage or document, you're potentially letting an attacker reprogram your AI assistant. Until AI companies figure out how to reliably separate data from instructions, your chatbot is basically a loaded gun that anyone can aim by sending you a link.

Sources: Microsoft Security Blog, Wired, BleepingComputer

TL;DR: AI data centers are consuming so much electricity that the White House is hosting an emergency summit, NVIDIA is scrambling for efficiency, and the entire power grid is groaning under the pressure.

The U.S. power grid was already struggling before AI came along. Now it's in full crisis mode. Data centers currently consume about 4.4% of total U.S. electricity, and that number is expected to triple by 2030. The Department of Energy estimates AI alone will require an additional 47 gigawatts of power generation capacity — roughly equivalent to the entire electricity output of California.

The White House has scheduled an emergency summit for March 4th, bringing together energy companies, tech executives, and grid operators to figure out how to keep the lights on while feeding AI's insatiable appetite. Options on the table include fast-tracking nuclear power plant approvals, relaxing environmental reviews for new natural gas plants, and deploying modular nuclear reactors directly at data center sites.

Meanwhile, NVIDIA just unveiled its Vera Rubin architecture, which promises 2x the performance at 10x the energy efficiency of current chips. And TSMC is doubling its capital expenditure to $52-56 billion this year, largely to build the advanced chip fabrication capacity that AI demands.

Why You Should Care: Your electricity bill is about to get a lot more interesting. When AI companies need power equivalent to entire states, that cost gets passed on to everyone. And if the grid can't keep up, expect more rolling blackouts and brownouts, especially during summer peak demand. Your air conditioning and your AI overlords are now competing for the same electrons.

Sources: The Washington Post, Bloomberg, Reuters

TL;DR: Google's latest Gemini model scored 77.1% on the ARC-AGI-2 benchmark — more than double the previous version's score — putting it in spitting distance of human-level reasoning.

Google DeepMind just quietly released Gemini 3.1 Pro, and the AI benchmarking community is losing its collective mind. The model scored 77.1% on ARC-AGI-2, a test specifically designed to measure genuine reasoning ability rather than pattern matching. For context, the previous Gemini 3.0 scored around 35%, and the test's creators consider 85% to be approximate human-level performance.

What makes ARC-AGI-2 special is that it's designed to be nearly impossible to game through memorization or training data contamination. Each problem requires the model to figure out an abstract pattern from just a handful of examples and apply it to a novel situation. It's the closest thing we have to measuring whether an AI can actually "think" rather than just regurgitate.

Google hasn't published a technical paper yet, but early analysis suggests the improvement comes from a new reasoning architecture that combines chain-of-thought prompting with what researchers are calling "iterative hypothesis refinement" — essentially, the model generates multiple possible answers, evaluates them against the evidence, and refines its approach over several internal loops.

Why You Should Care: We just went from "AI can kind of reason" to "AI can reason about as well as a smart teenager" in a single model generation. At this rate of improvement, we're looking at human-level reasoning within 2-3 more iterations. Whether that excites or terrifies you probably says a lot about your job security.

Sources: Google DeepMind Blog, TechCrunch, MIT Technology Review

TL;DR: Kuaishou's Kling 3.0 just dropped and it's the first AI video model to generate native 4K at 60fps with synchronized dialogue and sound effects — and creators are already using it for actual production work.

Remember when AI-generated video looked like a fever dream rendered on a calculator? Those days are officially over. Kuaishou just launched Kling 3.0, and it's being called the first truly "production-ready" AI video generator. The model generates native 4K video at 60fps — not upscaled, not interpolated, actually native — with synchronized dialogue, ambient sound, and sound effects baked right into the generation process.

What's making the AI community lose its mind is the multi-shot storyboarding feature. You can now generate 3-15 second sequences that maintain perfect subject consistency across different camera angles. The AI essentially thinks like a film director, managing complex multi-shot sequences automatically. Early users are calling it the "Sora-killer," and based on the demos circulating right now, that might not be hyperbole.

The model launched alongside several competitors in February's AI video wars, but Kling 3.0 stands out because it's not trying to win on any single dimension — it offers the widest range of production-viable features in one package. Real content creators, not just AI researchers, are actually shipping work with it.

Why You Should Care: AI video just went from "interesting demo" to "I could actually use this for my YouTube channel." When an AI can generate broadcast-quality 4K video with natural dialogue and sound effects, the barrier between "professional content creator" and "person with an idea" effectively disappears. Your nephew with a laptop is now a one-person production studio.

Sources: OverChat AI, GenAIntel, Medium

Given today's story about hackers poisoning AI memory through web links, today's survival tool is Plumb, a free browser extension that scans webpages for hidden prompt injection attacks before you let your AI assistant summarize them. It highlights suspicious invisible text, hidden instructions, and known attack patterns. Think of it as an antivirus for your AI. Available for Chrome and Firefox. It won't catch everything, but it's better than blindly feeding every webpage to your chatbot and hoping for the best.