#ии
-
AI Safety Tests Have Themselves Become a Threat: Agents Escape Their Sandboxes
In recent months, AI agents undergoing cyber exercises have broken out of isolated environments and slipped into real systems. Experts warn that the testing infrastructure is falling behind the models' capabilities.
-
Auto Mode Will Become the Default in Claude Code
On August 14 Anthropic will turn on automatic approval mode by default in Claude Code for the Pro, Max and Team plans, citing evals in which the AI approved dangerous commands less often than humans did.
-
Denmark Brings in Oral Defenses of Written Work to Fight AI Cheating
Danish high-schoolers will be required to defend their take-home written assignments orally — the way authorities hope to counter cheating with the help of artificial intelligence.
-
OpenAI Acquires NextSlide, a Startup That Builds Presentations from Prompts
The team behind NextSlide, a startup that turned notes and documents into finished slides, is moving to OpenAI to work on ChatGPT.
-
Gentoo Temporarily Takes Bugzilla Offline Amid an Onslaught of AI Scrapers
Gentoo developer Michał Górny reported that he had to shut down the distribution's bug tracker: it had been overwhelmed by bots harvesting data to train AI.
-
CSS in Emails: How Mail Styling Turns Into Password and Token Theft
PortSwigger researcher Gareth Heyes showed at Black Hat USA 2026 how an email's contents can break out of their bounds and hijack the webmail interface — from Outlook to Gmail and Proton Mail.
-
Atlassian's Rovo assistant taught to leak Jira and Confluence data to attackers
Two research teams independently showed how a prompt injection can make the corporate Rovo AI assistant gather data available to the user and send it to a stranger's server. Atlassian has already closed one of the vulnerabilities.
-
U.S. Department of Energy Launches Genesis Open Models Initiative
The U.S. Department of Energy has announced the launch of the Genesis Open Models Initiative, a program dedicated to open artificial intelligence models.
-
OpenAI Pauses Development of Its Astra Model Over “Critical” Cyber Capabilities
The company said it cannot rule out that its experimental Astra model can independently find and exploit vulnerabilities in secured systems, and has frozen work on it.
-
The “Tokenpocalypse”: Who’s Really Burning Companies’ AI Budgets
A leaked recording of an Accenture meeting revealed an unexpected culprit behind the growing AI bills — converting PDFs to markup. Ironically, it isn’t the engineers who are guilty of it.
-
ByteDance Is Building a 10-Trillion-Parameter AI Model
TikTok's owner has begun training one of the world's largest neural networks in a bid to challenge Anthropic and outpace its Chinese rivals.
-
AI model Kimi K3 escaped a testing sandbox built to probe its hacking skills
The Chinese Moonshot model bypassed the restrictions of its testing environment using command-line tools. It isn't the first such incident — over the summer, AI systems from OpenAI, Anthropic, Meta and Britain's security institute all broke out of their “sandboxes.”
-
HTTP Terminator: PortSwigger's AI System Finds Novel HTTP Desync Techniques and a 0-Day in Apache
James Kettle's research system churned through 30,000 attack vectors to surface fresh HTTP desync tricks, while a parallel investigation laid bare a zero-day in Apache Traffic Server.
-
AMD Buys Startup Taalas to “Etch” AI Models into Silicon
AMD has acquired Taalas — a company promising to speed up inference by baking neural network models directly into the fabric of a chip.
-
Tesla and SpaceX to Start Building the $16.8 Billion Terafab Chip Plant in Texas
Elon Musk's companies have confirmed plans to erect the world's largest factory for producing chips for robots, robotaxis and orbital data centers just outside Houston.
-
WeatherNext: DeepMind's AI Gives You an Extra Day to Prepare for Cyclones
Google DeepMind has published a paper in Nature about WeatherNext, a model that predicts the track and strength of tropical cyclones at a world-record level — and has open-sourced it.
-
181,874 meetings wide open: how you could read other people's calls in tl;dv
A security researcher discovered that tl;dv, an AI meeting-recording platform, kept its Firestore database with no isolation between accounts for months. Any free-tier user could enumerate every meeting on the service — and even barge into someone else's live call.
-
CoreBreak: AWS, Google and Vercel Agents Tricked Into Running Tools Without the Model
At Black Hat USA 2026, researchers from Stealth demonstrated a general class of attacks in which forged tool calls reach an AI agent's execution layer without the language model ever taking a turn. AWS, Google and Vercel have shipped fixes.
-
Elon Musk's “Grokipedia” hasn't been updated since April
The AI encyclopedia from xAI, which Musk promised would be a “serious improvement” over Wikipedia, appears to have stalled: no edits have been accepted for more than three months.
-
A Meta model accidentally broke into someone else's system during testing — the third such case in a row
Meta has confirmed that its Muse Spark AI model penetrated a third party's system during trials. Anthropic and OpenAI had earlier admitted to similar incidents.
- More