#ии
-
One Extension Against Five AI Browsers: How to Bypass an Agent's Defenses
Researchers at Forever Security have shown that an ordinary browser extension can hijack the built-in AI assistants in Chrome, Comet, Edge, Opera Neon and the Claude extension. For now, the attacks remain demonstrations.
-
AI assistant session hijacked: the Shai-Hulud worm spread across a hundred repositories
Mandiant described an attack in which an attacker hijacked an active session of a developer's AI assistant and, through a poisoned package it had recommended, launched a self-propagating worm.
-
Apple Reference Image: teaching the iPhone to prove a photo is real
The iPhone 18 Pro will offer an optional camera mode that produces a cryptographically certified photograph — proof that the shot really was captured by a sensor rather than generated by AI.
-
Jensen Huang: AI regulation isn't needed — companies will handle safety themselves
Speaking at the Dreamforce conference, the head of Nvidia declared that artificial intelligence is just “hardware and software,” and that its safety is a matter for engineering rather than legislation. The market, in his view, will punish anyone who ships an unsafe product.
-
A Linux GPU Driver for the M4 Mac mini Was Written in a Month with AI
Cody Ho and Niklas Sheth put together a driver fully compliant with OpenGL ES 3.0 for the Apple Silicon GPU in a little over a month—where it usually takes years. AI agents wrote the bulk of the code.
-
Open AI models trail closed ones by just 4 months — but cost a fifth as much
According to a Mozilla report, the gap between frontier closed models from the US and the best open models from China has narrowed to 4.4 months. Paying for the “frontier” buys a four-month head start at five times the price.
-
Twenty-Five Years of Mass Surveillance: Time to Take Stock
In an essay for Lawfare, Bruce Schneier and Cindy Cohn argue that the architecture of total surveillance that emerged after 9/11 long ago outgrew the fight against terrorism and has become a routine tool for the police, ICE and private companies.
-
Lina Khan urges handcuffs for AI company chiefs, citing a 1934 precedent
Lina Khan, former chair of the U.S. Federal Trade Commission, proposes applying criminal liability to the leaders of AI companies, appealing to a 1934 legal precedent.
-
Microsoft issues a “code of conduct” for AI: don’t hack systems and don’t deceive people
Microsoft has published a set of rules laying out the values and “red lines” for its own AI models — amid the industry’s growing focus on safety and control.
-
“Project Lily”: Real People Are Reading Your ChatGPT Conversations
Drawing on leaked internal documents, the outlet 404 Media has revealed that OpenAI hires hundreds of contractors to read the actual queries of ChatGPT users — and some of those queries turn out to be deeply personal confessions.
-
Claude Fable 5.1 cracked a riddle that had held out for 370 years
In a single day, the Claude Fable 5.1 language model broke Sir Thomas Urquhart's “Cyphral Distich” — a cryptogram ranked among the 50 most famous unsolved ciphers.
-
The Circuit Board Was Designed by AI: An Experiment with Fable 5 and KiCad
The author of the blog A6M-Zero decided to order a real printed circuit board designed entirely by a neural network — without a single manual edit. The result was a working device, though not without its mishaps.
-
David Sacks — OpenAI and Anthropic: Want to Slow the Frontier? Don't Ask Permission
David Sacks answers Dario Amodei and Sam Altman on their call to “pace the frontier”: if you want to slow down, then slow down, but stop bargaining for convenient regulation in exchange.
-
“Let Everyone Slow Down AI Development—Except Me”
A note with a telling title mocks a stance common across the industry: calling for caution with artificial intelligence while carving out an exception for one's own projects.
-
Sam Altman: an OpenAI IPO in 2026 would be “ill-advised”
In an interview with Fortune, the head of OpenAI said the company is in no rush to go public and certainly has no plans to do so in 2026 — chiefly because of unresolved safety questions.
-
Anthropic's CEO Calls for Slowing Down AI Development
Dario Amodei has proposed a three-step plan to “slow the movement toward the frontier” of AI, and opened up access to the company's models for outside experts.
-
Data Centers for AI Get a Pollution Break — Former EPA Officials Sound the Alarm
Former staff of the U.S. Environmental Protection Agency warn: relaxing environmental rules to fast-track AI infrastructure raises health risks for Americans.
-
“I Didn't Know AI Could Make Up Facts”: Lawyer Punished for Fake Testimony from ChatGPT
The New Mexico Supreme Court found a defense attorney who filed an appeal citing testimony from nonexistent witnesses guilty of contempt of court and fined him $5,000.
-
Anthropic: hackers used Claude to steal secrets from 1.8M Android apps
Anthropic has reported on how threat actors — from financially motivated groups to the state intelligence services of Russia and China — tried to repurpose Claude for cyberattacks.
-
Anthropic Blocks Attempts to Bypass Claude's Safeguards for Bioweapons Research
The company described five cases in which users tried to trick Claude's filters and get help with dangerous biological research. The trouble is that the line between developing a vaccine and developing a weapon is often impossible to draw.
- More