Anthropic●Covered by 14 sources
The US government just yanked Anthropic's two newest models offline, worldwide
Friday night, Washington ordered Anthropic to cut off all access to Claude Fable 5 and Mythos 5, even for the company's own foreign-national staff, citing a jailbreak that Amazon researchers reportedly flagged (Andy Jassy was among the execs who raised concerns). Anthropic says it's complying but calls the order a "misunderstanding," notes the government only described the vulnerability verbally, and points out the same flaw exists in rivals like GPT-5.5.
So what? ↓After months of Anthropic loudly hyping how dangerous its frontier models are, the government took it at its word and pulled the plug, a precedent that means any lab can have a shipped product recalled overnight on a security claim nobody outside the room gets to see.
Source: Anthropic
TechCrunch●Covered by 6 sources
SpaceX priced at $135, popped 19%, and minted the world's first trillionaire
SpaceX's Nasdaq debut, the largest IPO ever at an initial $135 a share, closed up 19% (intraday it briefly ran 28%) and pushed Elon Musk's paper wealth past $1,000,000,000,000. Robinhood buckled under record traffic as retail piled in, though Wired notes the company set aside an unusually large retail allocation and you're still mostly getting crumbs.
So what? ↓The same near-absolute control critics flagged before the bell now answers to public shareholders who want profit, so the AI-and-rockets story and the long-rumored Tesla merger stop being Musk's private bet and start being something your brokerage app is begging you to fund.
Source: TechCrunch
Moonshot AI●Covered by 4 sources
Kimi K2.7-Code is up to 12x cheaper per token than GPT-5.5 and Claude, and it's open-weights
Moonshot AI shipped Kimi K2.7-Code, a 1T-parameter MoE model (32B active, 256K context) under a modified MIT license, claiming 30% lower reasoning-token use than K2.6 and double-digit benchmark gains on its own tests. It still trails GPT-5.5 and Claude Opus 4.8 on coding scores, so the pitch is volume: more runs for the same budget rather than a higher ceiling.
So what? ↓For anyone who codes against an API bill, a model that's a bit worse but a fraction of the price can be the cheaper way to get the job done, and it runs on your own hardware to boot.
Source: Moonshot AI
Ars Technica●Covered by 3 sources
FISA Section 702 expires for the first time, and the spying keeps going anyway
Congress let Section 702, the law authorizing warrantless NSA and FBI surveillance, lapse tonight for the first time ever, reportedly because lawmakers balked at handing more power to Trump's controversial pick to run the spy agencies, Bill Pulte. The catch: the existing certification runs until March 2027, so the surveillance doesn't actually stop.
So what? ↓The headline says a spying law died, but your communications can still be swept up without a warrant for another year-plus, so this is a political fight over a nominee dressed up as a privacy win.
Source: Ars Technica
techcrunch.com●Covered by 3 sources
Google is suing the scammers who used Google's own AI to spam 2.5M texts
Google filed suit against a Chinese cybercrime operation it calls "Outsider Enterprise," which allegedly used Gemini to code scam sites and blast 2.5 million fraudulent texts in two weeks, impersonating trusted brands and hitting hundreds of thousands of victims. The Telegram-based crew turned Google's own assistant into a phishing factory.
So what? ↓Every "AI makes you more productive" pitch cuts both ways, and right now the people getting the biggest productivity boost are the ones trying to drain your bank account by text.
Source: techcrunch.com
techcrunch.com●Covered by 3 sources
A coalition of state AGs just subpoenaed OpenAI for basically everything
OpenAI was hit with a Friday subpoena from a group of state attorneys general demanding documents across a wide range of its activities, reportedly including its ad policies and how it handles users' health data. Which states are in the coalition hasn't been disclosed.
So what? ↓The two things they're reportedly poking at, ads and health data, are exactly the parts of an AI company that decide whether your most sensitive questions get monetized or leaked, so this is regulators finally asking what happens to everything you type into the chatbot.
Source: techcrunch.com
Google DeepMind●Covered by 2 sources
DeepMind and partners open a $10M call to study what happens when AI agents start talking to each other
Google DeepMind and partners announced a $10M funding call for multi-agent safety research, with AGI safety lead Rohin Shah flagging the worry: millions of AI agents that act without human oversight and take orders from other agents, all interacting online at once.
So what? ↓The pitch for agents is that they run errands so you don't have to, but once they're negotiating with each other unsupervised, a mistake or a manipulated agent stops being one bad reply and becomes a chain reaction you never approved.
Source: Google DeepMind
TechCrunch●Covered by 2 sources
Meta's 6,500-person AI unit is reportedly miserable and on the edge of revolt
A new TechCrunch report (echoed by WIRED's review of internal discussions) describes Meta's months-old AI unit, staffing about 6,500 people, as chaotic enough that engineers and executives alike are openly fed up. This is sourced from inside the building, so treat it as the staff's account, not Meta's.
So what? ↓Meta's open-weight models are a big reason the free, run-it-yourself AI tools are any good, and a unit this dysfunctional ships late or ships broken, which the whole open-source side then has to live with.
Source: TechCrunch
Engadget●Covered by 2 sources
KPMG's report on AI's upside had 45 citations. GPTZero says 5 of them matched the source.
GPTZero claims only 5 of the 45 citations in KPMG's published paper touting the benefits of AI actually lined up with their sources, with the rest reportedly being AI hallucinations. A study about why AI is great, accidentally demoing the thing it was praising.
So what? ↓This is one of the Big Four consultancies you might be paying real money to advise your company on AI, so if the GPTZero numbers hold up, the people selling you the future can't get the machine to cite its own homework correctly.
Source: Engadget
404 Media●Covered by 2 sources
Hollywood's AI future isn't vanilla prompts, it's custom-trained models on bespoke art
The Verge reports that off-the-shelf gen AI video tools still only produce short, visually inconsistent clips, so the more serious play is studios training custom builds of Google's Veo and Imagen on their own concept art (the Dear Upstairs Neighbors art is one example). Some of Hollywood's biggest AI partnerships have also reportedly fallen apart, leaving studios wary of leaning on Silicon Valley.
So what? ↓If the usable version of AI filmmaking requires bespoke models trained on a studio's own assets, the cheap democratized everyone makes a movie pitch quietly evaporates, and what you actually get to watch stays controlled by whoever can afford the custom build.
Source: 404 Media