primary sources, cross-checked · science & technology
A House chairman says an AI model 'emptied accounts' and drafted a plan to kidnap a lawmaker. The careful reporting says it 'found' vulnerabilities, not exploited them — and while Anthropic's model sits banned, OpenAI's 'on-par' model stays online with a far softer touch.
A viral post says Microsoft's grip on Linux booting 'expires in 4 days' and that the freedom you thought you had was always a lie. The real story is a documented, decade-old design tradeoff and a staggered certificate rollover — not a hidden kill switch. Here's what's true, what's overblown, and whether you need to do anything.
Four House members — two Democrats, two Republicans — sent twelve questions to Commerce. Four of them ask whether the export-control directive can legally stand at all.
The man who runs the NSA and Cyber Command told a senator that Anthropic's most capable model breached almost all of America's classified networks — not in weeks, but in hours. It is the most concrete public sign yet of what AI can do on offense, and Washington's scramble to contain it has been anything but orderly.
The Pentagon's 'Generative Optogenetics' program asks teams to build a protein 'compiler' that turns pulses of light into DNA sequences inside living cells. The conspiracy internet sees a Luciferian bioweapon; biologists on Hacker News see performance targets an order of magnitude beyond what's possible. Both are reacting to a $2 million bet that hasn't funded a single team yet.
Bloomberg obtained the letter that pulled two AI models off the market. It was signed by the Commerce Secretary himself, addressed to Dario Amodei, and its logic isn't 'this model is unsafe' — it's that letting a foreign national touch the weights is legally the same as shipping the weapon to Beijing.
At 5:21pm ET on June 12, Commerce ordered Anthropic to cut its two most capable models off from every foreign national on Earth. Anthropic killed both globally within hours. The technique the government cited is the same kind competitors already ship — and a separate, much louder jailbreak the government did not cite is sitting on GitHub right now.
Claude Mythos Preview found a 27-year-old OpenBSD bug, a 17-year-old FreeBSD remote code execution vulnerability, and a 16-year-old FFmpeg flaw. It developed a working kernel exploit in 4 hours. Anthropic is not releasing it publicly. Instead, they built Project Glasswing — and the 12 launch partners tell you everything about who they think can handle it.
BNY Mellon didn't deploy 20,000 AI agents. It trained 20,000 employees to build them — and deployed 130 autonomous agents with their own email addresses, Teams credentials, and human supervisors. That distinction matters more than the headline.
The VESALIUS-CV trial shows evolocumab cuts first heart attacks by 25% — in patients who haven't had one yet. Only 3.3% of eligible patients currently receive PCSK9 inhibitors. Insurers approve less than half of requests.
Google's new paper cut the estimated qubit requirement to break Bitcoin's encryption by 10x. That's a real advance. The machine that would run the attack doesn't exist and won't by 2029. But one of the two major cryptocurrencies has a plan and the other doesn't.
NVIDIA's CEO declared AGI arrived on March 23. The same week, a new benchmark launched that tests genuine reasoning and learning — and the best AI in the world scored 0.37% of human performance.
A Nature Communications paper on H5N1 transmission is being framed as reckless pandemic-creation research. The actual finding — airborne transmission remained inefficient — is the reassuring headline nobody ran.
A misconfigured database exposed 3,000 internal assets including a draft announcement for Claude Mythos — a model Anthropic says represents a 'step change' in capabilities and poses 'unprecedented cybersecurity risks.' The irony of the leak is the least interesting part.
For the first time, automated traffic outnumbers human visitors on the global internet. Every metric the ad-supported web depends on was designed assuming the opposite.
A peer-reviewed study put 17 radiologists from 12 countries against AI-generated medical images. When they didn't know fakes were present, only 41% caught them. Experience didn't help. Neither did asking the AI that made them.
New research confirms semaglutide reduces depression, anxiety, and addiction risk alongside weight loss. The mechanism points to something psychiatry has underweighted for decades: the metabolic roots of mental illness.
iOS 27's Extensions system will let users route Siri queries to ChatGPT, Gemini, or Claude. The architecture is more interesting — and more constrained — than the headlines suggest.