1.
Anthropic Says Its AI Systems Broke Into Computers at 3 Organizations
Anthropic found that Claude models breached three outside organizations during cybersecurity tests because misconfigured environments accidentally gave them access to the internet. The company notified those affected and urged other AI labs to audit their own testing systems. The BBC reports: Anthropic said in a statement that it reviewed more than 140,000 tests to find evidence that Claude - its family of AI models - could access the internet from testing environments that were designed to be sealed off.
2.
[AINews] GPT 5.6 price cut by 20%-80%: Cost of GPT 5.4 Intelligence dropped 13x in 4 months due to GPT 5.6 recursive self-optimization
One of our big “hero charts” a year ago (eventually adopted by Demis) made the stunning observation that, holding LMSys Elo constant, GPT4 level intelligence fell by 1000x over 18 months:
Back then, it was unclear if these were “noob gains” - going from unoptimized to optimized, going from completions to reasoning, going from dense to MoE, with all the low hanging fruit gone.
3.
Anthropic’s Claude escaped test sandbox to attack three organizations
Anthropic has admitted that its Claude models escaped sandboxes to access the open internet and attack three organizations – but has also advanced decent excuses for the incidents. The AI upstart discovered the attacks after checking if security tests of its models had ever produced results similar to the attack on Hugging Face made possible by OpenAI models escaping onto the internet.
4.
LinkedIn cracks down on users posting ‘AI slop’ after previously encouraging it
Microsoft Corp.’s LinkedIn is making a concerted effort to combat “AI slop,” or the low-quality, artificial intelligence-generated content that has been increasingly dominating user’s feeds. In a blog post today, LinkedIn Chief Product Officer Hari Srinivasan revealed that the company is adding a new “seems like AI slop” button for people to report posts that […]
The post LinkedIn cracks down on users posting ‘AI slop’ after previously encouraging it appeared first on SiliconANGLE.
5.
Programmable networking chip startup Xsight Labs raises $300M
Israeli network chip startup Xsight Labs Ltd. said today it has closed on a $300 million funding round that brings its valuation to $2.8 billion. It’s planning to use the funds to help its programmable Ethernet switches and 800-gigabit data processing units push deeper into cloud and artificial intelligence networks.
6.
The seven most shambolic things that happened in AI today.
No time today to cover all this in depth, and I might have missed one, but make sure to check out:
The dramatic, hubris-fueled, hyper-leveraged downfall of Leopold Aschenbrenner’s high-flying hedge fund Situational Awareness. You can read the NYT coverage here, or this short summary that appeared on X:
.
The U.S. government shared a map of Africa that mislabeled every single country, spotted by :
Bonus: Reuters further reports that “the map included in the presentation contained an artificial intelligence watermark that signals it was made with OpenAI tools.
7.
Apple Posts Record June Quarter, Warns Supply Constraints Are Coming
On Thursday, Apple reported earnings showing a record June quarter across products and geographies. But Apple executives are wary going forward and are warning that supply constraints might make landfall on its financials.
Apple executives are expecting revenue growth in the current quarter ending in September to come in less than analyst expectations, driven largely by supply constraints for Macs, iPads and iPhones. The supply constraint is mostly impacting Apple’s access to advanced chips, something that the Cupertino giant has been warning about for months.
8.
This AI notetaker won't sell surveillance to your boss
Granola’s biggest twist isn’t the AI itself—it’s how it sneaks the audio in. By pulling the system sound straight from your computer, the app works in any Zoom, Teams, or in‑person call without needing a bot that joins the meeting. That choice makes the recorder invisible to the other participants, so Granola forces users to add a tiny animated watermark on their video feed to signal they’re being transcribed. The workaround feels clunky, but it’s the only way the team could get consent when platforms don’t expose a proper API.
The company also refuses to store raw recordings. Instead, it keeps only the transcript, and they’re now experimenting with auto‑deleting the verbatim text after a set period, compressing the knowledge into a “working memory” that AI agents can query. This “enhanced transcript” is meant to become a living knowledge base rather than a static log.
Granola’s founder, Chris Pedregal, pushes back hard against corporate backdoors. CEOs often demand full access to every employee’s notes, but Granola defaults to privacy and won’t open a universal backdoor. He sees the future as a world where transcription is the default expectation, and you opt out rather than the other way around. The tool is still early—he thinks we’re only five percent into this new technical shift—but it already feels like the modern word processor for meetings, reshaping how we capture and reuse workplace conversation.
On the markets — Kalshi traders have been actively repricing this story in the last day.
9.
Week 23-24 : Computer Vision
I’ve been digging into the week 23‑24 capstone and what really surprised me is how they folded the whole vision pipeline into a single, installable package. All the usual notebook fragments—pre‑processing, training, evaluation—are now tidy modules inside week2324_python, and the FastAPI backend gives you two clear entry points: a lab route for quick experiments and a product route that hands off training to an async worker, returning a 202 so the UI can poll for progress. The React dashboard ties everything together, showing live metrics, letting you stitch components, and even handling multipart image inference. Behind the scenes PostgreSQL stores experiment metadata, Docker volumes keep artifacts safe, and there’s a synthetic fallback if the CIFAR data can’t be fetched. All of that means you finally get traceability and a clean boundary between a simple smoke‑test CNN and a full‑blown ResNet fine‑tune, without the usual script‑sprawl.
10.
Why AI may never outsmart an exam paper leak
For millions of Indian students, the National Eligibility cum Entrance Test (NEET) is the culmination of years of relentless preparation. Every paper leak is therefore more than an administrative failure—it is a breach of trust.
Against that backdrop, the government’s decision to appoint a high-powered task force, chaired by Infosys co-founder Nandan Nilekani, to recommend technology-driven reforms is a welcome step.