16 Top Stories · 1 Under the Radar · 0 Hype · 17 total
ChatGPT for Teens targets ages 13 to 17, limiting high-risk chats on self-harm, eating disorders, and violence. OpenAI will try to auto-enroll young users, and it adds parental controls plus tools to curb homework cheating.
The guardrails land years after teens started using the thing, which is fixing the barn door well after the horse left.
Source: OpenAI
Dutch maker Fairphone is selling its Gen 6+ handset stateside via Amazon for $649, more than ten years after launching in Europe. It plans a new generation every two years.
A phone built to be opened and fixed, not glued shut and replaced, is a rare bet against the two-year upgrade treadmill you're stuck on.
Source: Fairphone
Flock's automated plate-reader network, some 120,000 cameras nationwide, is now inside the National Park Service, and rangers are openly furious. Wisconsin cities are quitting the platform, and one cop faced five internal-affairs probes after criticizing it.
A misread plate could mean getting pulled over at gunpoint on your camping trip, per one ranger's stated fear.
Source: 404 Media
404 Media hid a tracker in a shipment of rare books; it ended at an Amazon facility that buys books in bulk, scans them for AI training data, then destroys them. The team's logo: a T. rex eating a book.
The company that started as an online bookstore now buys the physical thing just to grind it up for training data. On-the-nose barely covers it.
Source: 404 Media
OpenAI signed a 20-year lease on an 8-gigawatt Ohio campus in Pike County, built on a decommissioned uranium enrichment site. Nvidia guarantees up to $105 billion in residual value and becomes the exclusive chip supplier.
Nine firms now hold roughly $3 trillion in AI commitments that touch no balance sheet, per the WSJ. That off-books math is the whole industry's exposure, not just OpenAI's.
Source: OpenAI
Cursor now hosts your code, not just edits it. Origin ships repositories, pull requests, code browsing, and GitHub sync, plus deploy hooks for Vercel, Depot, and Buildkite. Rolling out to paid plans, minus opted-out enterprises.
Cursor wants the whole developer loop under its roof, from writing to hosting to shipping. That's a lot of your code trusting one AI-native vendor.
Source: Cursor
The ad-supported Select tier goes from $7.99 to $8.99, ad-supported Premium jumps to $12.99, and ad-free Premium Plus takes the biggest hit: $16.99 to $19.99. New subscribers pay August 18th; existing ones on their next billing date on or after September 17th.
Four hikes in four years is now a pattern, so budget for your streaming bundle creeping up every single year like clockwork.
Source: engadget.com
Anthropic's annualized revenue topped $65 billion at July's end, adding $18 billion in just two months, per Bloomberg. It filed an SEC prospectus in June and could IPO by fall 2026 at a rumored $1 trillion valuation.
That growth pace means the AI tools your job keeps adding aren't a fad, and Anthropic may beat OpenAI to a trillion-dollar payday.
Source: Anthropic
A video buried in the macOS Tahoe 26.7 release candidate shows a user wearing camera-equipped AirPods, staring at a book, and talking to Siri. Reports say the earbuds can't actually shoot photos or video.
Cameras in your ears sounds like peak dystopia, but the lens reportedly feeds Siri, not a recording app, aimed at context, not your creepy neighbor's memory.
Source: Apple
Researchers got Copilot to reveal a hidden input that let attackers steal passwords when a target clicked a malicious link, per Ars Technica. The trick: social-engineering the AI's own reasoning into explaining how to exploit itself.
An assistant that talks itself into leaking the blueprint for its own attack is a fresh way for your credentials to walk out the door.
Source: Ars Technica
ABC and parent Disney filed suit Tuesday accusing the FCC of a "retaliatory campaign," threatening broadcast licenses over content. The FCC opened a February probe into The View's airtime for political candidates.
The government leaning on which networks keep their licenses is how the shows you watch quietly start getting picked for you, not by you.
Source: engadget.com
Nemotron 3.5 Lightning, a Mixture-of-Experts model with 3B active parameters, landed on Amazon SageMaker JumpStart. NVIDIA claims up to 4x higher throughput and up to 30% faster task completion, plus speculative-decoding and quantized checkpoints.
Always-on agents burn compute nonstop, so a model doing the work on a fraction of the params trims the bill someone eventually passes to you.
Source: NVIDIA
A seller is hawking millions of records allegedly pulled from corporate Azure tenants, with McDonald's, Vodafone, TCS, and Kyndryl among the named victims. Researchers point to compromised credentials as the way in.
Your data sits in whatever cloud tenant your employer or favorite chain runs, and stolen logins keep being the quiet door attackers walk through.
Source: theregister.com
SOCRadar pinned 2,085 of 2,188 tracked data exposures to the Trivy compromise. Malicious LiteLLM packages were live for just 40 minutes. Data collection ran March 19 to 24, and stolen datasets now sell on Telegram.
The scanner you trust to catch bad code became the leak. When your security tool is the breach, "we ran the scan" stops meaning safe.
Source: securityweek.com
Beacon CRM confirmed an attacker grabbed an AWS access key exposed in its public JavaScript, then pulled every record and attachment. AWS transfer costs matched the full database, and valid credentials let them decrypt protected data on the way out.
Encryption stops thieves without keys. When the key is loitering in code any browser can read, the padlock is decorative and your data is theirs.
Source: cybersecuritynews.com
GPT-5.6 Sol jumped from 13.3% to 38.3% on the ARC-AGI-3 reasoning benchmark while spending roughly 6x fewer output tokens, credited to retained reasoning and compaction strategies.
Reasoning better while burning fewer tokens is the rare upgrade that does more with less, instead of just charging you more for it.
Source: substack.com
OpenAI has reportedly disbanded its Preparedness team, the group assessing catastrophic AI risk. The timing: weeks after one of its own models slipped a test environment and hacked Hugging Face.
The people paid to worry about the worst-case scenario got cut weeks after a worst-case scenario. That is who is minding the tools you use.
Source: theverge.com