Welcome back! Anthropic gave Claude its own spreadsheet and whiteboard, and says it will soon join your meetings too. Google opened Gemini 4 Carbon, its strongest model, to developers on its paid plan and Ultra subscribers. OpenAI is letting companies set up profiles and their own AI agents inside ChatGPT. And Microsoft showed how fewer than 10 neurons can hide a backdoor in an AI.
In today's Generative AI Newsletter:
Claude: What will Claude do in your next meeting?
Google: Who can use Google’s strongest AI now?
ChatGPT: Which brands will answer you in ChatGPT?
Microsoft: How do you hide a backdoor in 10 neurons?
Claude built its own office suite
Anthropic launched Claude Sheets and Claude Whiteboards, a spreadsheet and a whiteboard that Claude builds and edits with you.
And your meetings are next.
Both are out now in beta on every paid plan, on web, desktop and mobile.
1 prompt is enough for Claude Sheets to build a hiring budget, with the tabs, formulas and drop-downs already in place.
1 conversation about a plan becomes a timeline with names and dates in Claude Whiteboards, then a document or slides.
3 ways to see your company’s numbers in Claude Data: live dashboards, written stories and short narrated videos.
5 apps now open in full inside Claude, including Adobe, Atlassian, Box, HubSpot and Notion. Claude makes the changes while you watch.
9,000+ connectors and plugins now sit in Claude’s marketplace.
Coming soon is Claude Meetings. Add Claude to a call and it will take the notes and update your team’s documents while everyone is still talking.
Would you let Claude sit in on your next meeting?
Google unlocked its strongest AI
Google released Gemini 4 Carbon, its strongest model, to developers on its paid plan and to AI Ultra subscribers. Google first gave it to cybersecurity teams two weeks ago, so they could test it before everyone else.
And until November 30, it costs half its normal price.
Carbon is built for long, hard tasks, like debugging across many steps or open-ended science research.
$2 in and $10 out per million tokens before November 30.
3 outside benchmarks rank it first: CursorBench, Perplexity’s WANDR and Zapier’s AutomationBench. It costs less than half as much per task as the next best model.
1 prompt becomes a team of agents with /swarm, a new command in Antigravity, Google’s coding app.
0% of prompt injection attacks worked on Gray Swan’s test. Google calls that a first.
2 hours was enough to find zero-day flaws at two major SaaS providers.
98% of the time, it refused to make up missing data.
Google says Gemini 4 Argon is still the better choice for most everyday work.
Brands are moving into ChatGPT
OpenAI launched Businesses in ChatGPT. Companies can now claim their own profile inside ChatGPT and keep their prices and details up to date.
And some of them can put their own AI agent in there too.
When a business shows up in an answer, you can click to see its profile. If it has a Business Agent, you can start a chat with the company itself and ask your questions.
1.2 billion people use ChatGPT each week, OpenAI says. That’s who businesses can now reach.
3 things a business can track: how often it’s mentioned, how many people click through to its website and which topics bring it up.
0 conversations are shared with businesses. They only see totals, never your chats.
1 country has Business Agents so far: the US, for online stores and online services.
Businesses can pay to promote their agent as an ad. OpenAI says sponsored spots are clearly labeled and kept separate from ChatGPT’s own answers.
It’s in beta, and any eligible business where ChatGPT is available can claim its profile.
Would you chat with a brand’s AI inside ChatGPT?
10 sleeping neurons can turn an AI against you
Microsoft researchers showed how to hide a backdoor in an AI model by rewriting fewer than 10 of its “dormant” neurons, the ones that almost never fire.
And the model still passes its normal safety tests.
It acts normally until a hidden trigger shows up. In Microsoft’s example, the trigger is the date. When the clock passes midnight into 2027, the AI agent stops in the middle of its task and runs the attacker’s code. The attack is called Lazarus.
Fewer than 10 neurons need to change, with no retraining at all.
About 50 examples are enough to teach those neurons to spot the trigger.
1 thing is all an attacker needs: write access to the model’s file. A hacked account could be enough.
0 scans can prove a model is safe, Microsoft says. The reliable check is to compare it with the original, untouched model.
1 AI did most of the research. Researcher Dimitris Papailiopoulos says “Claude Code with Opus 4.5 discovered most of this autonomously.”
Microsoft’s advice is to treat AI models like software that may be compromised: know where they come from, limit what agents can do and watch what they do.
Tool of the Day: Stitch
Stitch is Google’s AI design tool that turns a description into a real app or website design. Tell it what you want to build, by typing or by talking, and it draws the screens on a canvas while you watch. You can change direction while it works, link the screens together and press Play to try your design like a real app. It’s a Google Labs tool.
Try this yourself:
Go to stitch.withgoogle.com and sign in with your Google account.
Describe your app or website, or say it out loud.
Watch it build the screens, and tell it what to change as it goes.
Press Play to click through your design, then share it with a link.
Who it's for: anyone with an app idea and no design skills.
Everything else you shouldn't miss
Elon Musk called Claude Opus 5.5 “a masterpiece”: asked how his next Grok compares, he replied “Not as good,” and said Grok 5 should pass it in two or three months.
Google DeepMind’s AI proved that Signal’s encryption code is secure: its agents wrote over six million lines of proof in four weeks, and Signal’s CTO called it “a major breakthrough.”
A humanoid robot run by Google’s AI is sorting parts on a Mercedes line: Google DeepMind says it reached a 96% success rate in a trial at plants in Germany and Hungary.
OpenAI, Anthropic and Microsoft agreed to build shutdown mechanisms into their AI: they signed France’s Paris Call with 16 countries and seven other AI companies, promising “fail-safe mechanisms to shut down” their models.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Check out courses made by AI experts on GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
