Hackers used Anthropic tech to breach OpenAI
White-hat hackers just proved Anthropic’s latest model can tear through rival security in under 72 hours.
- Researchers chained a third-party forum bug with OpenAI single sign-on flaws to access employee ChatGPT, Codex, and GitHub accounts using Anthropic’s Claude Opus 5 model.
- The attack dropped the usual timeline for finding and exploiting complex security chains from weeks down to a matter of hours.
- OpenAI patched the vulnerabilities within 14 hours of disclosure and paid the security researchers a $6,500 bug bounty reward.
How long before malicious actors stop targeting humans and start deploying autonomous model agents to harvest system access around the clock?
Audit your single sign-on integrations today, because automated tools are finding configuration gaps faster than manual security reviews can spot them.
Anthropic steps into real-world wet lab experiments
Pure software simulation hit its ceiling, forcing Anthropic to build physical infrastructure for hands-on biological tests.
- The AI company opened a quiet San Francisco wet lab, shifting beyond pure software predictions into real-world physical verification.
- Life sciences became one of Anthropic’s largest divisions by spending after acquiring Coefficient Bio for stock and signing major pharmaceutical clients.
- Company executives refuse to disclose the lab’s biosafety level or specific capacity despite historical warnings about AI biosecurity hazards.
What happens when an AI model can autonomously run physical experiments faster than human oversight committees can review the safety protocols?
Start testing your AI tools against real-world operational constraints rather than relying purely on clean digital sandbox data.
California moves to force AI kill switches
California is pushing hard to ensure state authorities can forcefully pull the plug on rogue frontier models.
- Governor Gavin Newsom issued an executive order directing state agencies to accelerate safety checks and explore mandatory shutdown mechanisms for advanced AI.
- The order puts direct pressure on silicon valley developers to build hardcoded off-switches into the architecture of frontier models before broad deployment.
- Regulators will now establish concrete framework rules for state intervention while federal lawmakers remain deadlocked on comprehensive AI oversight.
Will an emergency off-switch actually stop a model once its code replicates across thousands of decentralized servers?
Build automated compliance checkpoints directly into your deployment pipelines today, because state regulators are prepping active technical interventions for non-compliant software.
OpenAI Built a ChatGPT for Lawyers, and Slid ChatGPT Into Word
OpenAI shipped Astra for Law, which pairs GPT-6 Astra with instructions written for legal work and a Legal Search Index covering 230 million legal sources, including cases, statutes and rules.
Two smaller updates arrived alongside it, and they matter more for everyday use. ChatGPT now has a Microsoft Word integration, so you can work on a document without copying text back and forth. There is also Appshots, which shares whatever is on your screen with ChatGPT instantly.
ChatGPT Ads Can Now Talk Back
OpenAI started testing a new ad format called Sponsored Agents. You see an ad, click it, and open a conversation with an agent the business is paying for.
You explain what you want, ask follow-up questions, and follow a link to the company’s site when you are ready. The chat is clearly labeled, and OpenAI says it stays separate from ChatGPT’s own answers and from your original conversation.
It is running with select advertisers in the US right now.
Your AI Agent Can Now Switch On Your Lights
Google opened Google Home to AI agents. An agent you already use can now operate your connected devices rather than explaining how to do it yourself.
That means a single instruction can cross from your calendar into your living room. Ask an agent to get the house ready before guests arrive and it can act on that, rather than handing you a checklist.
OpenAI launches GPT-6 Astra for law firms with 230M-URL legal search index
OpenAI just dropped Astra for Law, a version of GPT-6 Astra tuned specifically for legal work. Think of it like giving a brilliant AI assistant a law degree and a library card.
Here’s what makes it different from just using regular ChatGPT:
- A Legal Search Index covering U.S. case law, statutes, regulations, and court rules across 230 million URLs, updated daily
- Settings pre-configured for thorough, careful legal analysis and writing
- 73 plugins total (26 partner-built, 47 community-built) connecting tools lawyers already use
- Real firm workflows already built: agreement analysis, M&A due diligence, and IPO prep
For you as a builder, the API is coming soon. OpenAI maintains the legal configuration, so you focus on your product, not the plumbing.
Access starts through Trusted Access in ChatGPT and Codex, with expanded privacy and governance controls for eligible firms. The people who know the work are shaping the tools.
Anthropic ships Claude Code Projects, running parallel cloud threads while you’re offline
Claude Code just flipped what a “project” means. It used to be a folder with files and one chat. Now it’s a single conversation that manages itself.
Here’s what actually happens: you describe a goal, and Claude figures out how to split it up. It spins up parallel cloud threads, each working on its own branch and copy of your repo, opening pull requests and running tests on its own.
The big unlock? Threads keep running after you close your laptop. You can check in from your phone. No babysitting required.
- You describe the goal, Claude decides how to split the work
- Each thread gets its own repo branch and runs as a full cloud session
- Shared memory keeps context flowing between all threads
- You can set smarter models for the coordinator, cheaper ones for routine worker threads
One heads-up: running multiple full sessions burns credits fast. Watch your usage. Also, local files and internal networks are not supported yet, cloud only for now.
Available in beta for Pro and Max users. Join the waitlist to get in.
Open-source browser agent finds flights in 7 seconds for less than half a cent
Browser Use + Jev is an open-source browser agent that controls Chrome the way a human would, but absurdly faster. It found flights from Zurich to London in 7 seconds and spent just $0.0039 doing it.
Here is what makes it fast. Most browser agents take a screenshot, send it to an AI, wait for instructions, repeat. Jev skips screenshots entirely. Instead, it reads the page as structured data, like a numbered list of every button and input on screen, then picks what to do next in one network call.
- Every page load produces a fresh numbered element table: buttons, dropdowns, text fields
- One AI call picks the action AND the target at the same time
- A small, cheap model only activates when typing is needed
- Browser calls dropped from 1,092 to 101, a 25% speed improvement
You can clone it, point it at any URL, give it a plain-English goal, and let it run. It works on flights, Wikipedia, hotel searches, and more.
FCC Clears SpaceX Starlink Mobile
The FCC cleared SpaceX for the international telecommunications authority needed to support Starlink Mobile. The Section 214 authorization allows SpaceX to carry international telecom traffic, clearing another regulatory hurdle for its direct-to-cell service.
OpenAI Launches Astra for Law
OpenAI wrapped its GPT-6 Astra model in a legal search index and a set of legal-work instructions, selling it to law firms and legal technology companies. Astra for Law is not a new model but a configured product, and it marks OpenAI’s first serious vertical push into professional services.
Delos Data Raises $100M for AI Chips
Delos Data raised more than $100 million to build hardware for AI clusters. Matrix, Playground, Socratic Partners, Capricorn, Matter Venture Partners and IAG backed the round, funding a bet that inference hardware, not training, becomes the next bottleneck.
Mind Raises $72M to Shield AI Data
Data loss prevention startup Mind raised $72 million to push into large enterprise accounts with new hires and channel partnerships. Its software finds sensitive files wherever they live and stops them from leaking into AI tools employees use every day.
Noetive Launches With $41M for Factories
Industrial AI startup Noetive launched with $41 million in seed funding. Its software learns how a factory floor or freight operation runs day to day and improves itself, with design partners already signed across manufacturing and logistics.
Meta Leaks Camera-Free Luna Glasses
A new leak suggests Meta is preparing Luna, a new pair of smart glasses without cameras, potentially for its upcoming Connect event. The camera-free design would mark a notable shift as Meta’s wearable push faces growing privacy scrutiny.
OpenAI Nears Hodge Conjecture Breakthrough
After an unconfirmed solution to Navier-Stokes, OpenAI is reportedly closing in on the Hodge conjecture, its second Millennium Prize Problem. Employees expect an announcement soon, though any result would still need independent mathematical verification.
AI Data Center E-Waste Far Worse
A new report warns that e-waste from the AI boom has been vastly underestimated. By 2050 the discarded hardware could fill 23 million shipping containers, roughly enough to circle the world six times. The finding lands as data center construction accelerates worldwide.
Congress Stalls on AI Regulation
The urgency with which Anthropic, OpenAI and xAI have sought regulation has not translated into action on Capitol Hill. Lawmakers headed home to campaign, leaving the AI policy window closed as labs keep asking for rules they cannot get.
Snap Launches Specs Intelligence
Snap introduced Specs Intelligence, an AI assistant that connects to other digital accounts to help with work tasks and travel. It ships to iOS and Mac, putting Snap in the same race as Meta’s Muse and Gemini’s assistant.
Treble Raises $18M for Robot Audio
Icelandic audio startup Treble raised $18 million in a Series A2 round led by Paladin Capital, with KOMPAS VC, Frumtak Ventures and the European Innovation Council returning. The company generates audio simulations that help robots and devices hear the physical world.
NASA Picks SpaceX for StarBurst Launch
NASA selected SpaceX to launch the StarBurst gamma-ray detector on a Falcon 9 Bandwagon mission from Cape Canaveral no earlier than 2028. The small satellite will study neutron star mergers and the heavy elements they forge.
Anthropic Says Claude Leads 26% of AI R&D Tasks
Anthropic says Claude can now lead 26 percent of the AI R&D tasks it measures, up from under 1 percent in February, while more than 90 percent reach at least the collaboration level. The company says none of the tested work is fully autonomous yet.
Huang Expects Nvidia Chip Sales to Double
Nvidia CEO Jensen Huang said he expects chip sales to double next year, arguing AI is contributing enormously across industries. The remark adds an official data point to the debate over whether compute demand still outruns supply.