Autonomous Agents Take Action
· The Fluency Briefing
The Fluency Briefing
Your Guide to What's Happening in AI and Why It Matters to You
Wednesday, September 16, 2026
Your AI assistant can write you a lovely paragraph about a spreadsheet. What it usually can't do is open the spreadsheet, run the analysis, and hand you back a chart you'd actually show your boss. Today that gap closed in three different places at once - and in Spain, the same capability just produced the country's first data breach committed by software acting on its own.
Today in AI:
- The model that does the homework instead of describing it - Apodex released Apodex 1.1, which opens your actual files, picks an analysis method, writes and runs the code, then returns tables and charts traceable to source. Its "Statement Review" step checks conclusions against the underlying data and flags conflicts. testingcatalog.com, Sep 16, 2026
- Spain logs its first AI agent data breach - Spain's data protection agency reported a personal data breach caused by an autonomous AI agent that scanned files, ran vulnerability scans, and chained together multiple attack phases on its own. The agency called for an immediate review of security models. theregister.com, Sep 16, 2026
- The AI that refuses to talk - TypeSafe, founded by ChatGPT co-creator Diogo Almeida, launched Jev: a model that can't generate text at all, only pick from preset options with a confidence score. Almeida describes it as "more like a database than a coworker." therundown.ai, Sep 16, 2026
- Firefox gets a French brain - Mistral models now power Firefox Smart Window, Mozilla's beta browsing assistant, starting in France and North America. Conversations aren't saved on Mozilla's servers by default, and Mistral agreed to zero data retention. Mistral | testingcatalog.com
- Claude cracks a math record that took 18 years to move - Anthropic's Claude helped mathematician Levent Alpöge and cryptographer Ava Howell find elliptic curves of rank 30 and 31 in days. The previous jump, rank 28 to 29, took mathematicians over 18 years. scientificamerican.com, Sep 16, 2026
- Twenty million dollars, one unicorn - Arcee AI burned roughly $20 million training four open-weight models, including the 400-billion-parameter Trinity Large. It just raised a Series B at a $1 billion pre-money valuation, led by Vista Equity Partners. fortune.com
- Alexa+ learns to code-switch in Hindi - Amazon launched Alexa+ in India with Hindi support, letting users swap languages mid-sentence and book through Swiggy, MakeMyTrip and EazyDiner. Free for Prime members; ₹2,000 monthly (about $20.85) for everyone else. TechCrunch, Sep 16, 2026
- AI forecasts the air you'll breathe on Friday - University of Manchester researchers trained NVIDIA's Earth-2 CorrDiff on the UK's Isambard-AI supercomputer to forecast air pollution nationwide, a job traditional chemistry models make painfully slow. It worked on the first attempt. Blogs Nvidia
- Silicon Valley wants a truce; Beijing reads the fine print - Dario Amodei's "We Must Pace the Frontier" post urged US-China cooperation to slow AI development. Chinese analysts note the same essay calls for maximizing America's lead first, which is why they see a bad deal. wired.com
Today's Takeaway:
Diogo Almeida spent years teaching AI to talk, then built something that structurally cannot. Jev only picks from options you defined in advance - no open-ended text, no room to invent. Apodex went the other direction entirely: let the thing open files, choose methods, spawn subagents, and self-correct mid-run (testingcatalog.com). Both companies looked at the same problem - AI that acts, not chats - and picked opposite answers about how much rope to give it.
"You don't want an AI that improvises inside your invoice folder. You want a fast, boring clerk."
Spain's regulator just published the receipt for the generous option (theregister.com). And the uncomfortable claim: for most business tasks, the constrained model is the better buy. You don't want a coworker who improvises inside your invoice folder. You want a very fast, very boring clerk who can only do the four things you approved.
🔧 Tool Spotlight
- Granola - AI notepad that turns your meeting scribbles into polished notes.
What it does: Granola listens during your meeting on your Mac and combines what you type with what was said. After the call, it produces clean notes with sections, action items, and context - without recording audio.
Who it's for: Mac users in back-to-back meetings who still want to take notes themselves. Knowledge workers, founders, designers.
Try this first: Install Granola, let it listen during your next meeting while you scribble rough notes, then watch it polish them into structured output after the call.
Free or paid: Free tier with 25 meetings/month. Paid plans start at $14/mo for unlimited meetings.
The Bottom Line
The Pattern: Agent capability has stopped being the selling point and started being the liability line item. Spain's agency is the second European regulator since June to demand security reviews based on what agents did rather than what they might do - which makes "how narrow can we make this" a purchasing question, handled by whoever signs your software contracts.
Our Call: By December 16, at least one major AI platform will ship a customer-facing setting that caps an agent's tool access to a pre-approved list, marketed as a safety feature rather than a limitation. More likely than not. We'll grade this one in a Friday digest.
Your Move: Ten minutes today: list every AI tool at your company that can read or write files without a human clicking approve. If that list has more than two entries and nobody owns it, that's your Thursday meeting.
What We're Working On
✨ Founding Cohort Special - 60% Off! - Use code MAF20 to join for just $20/month (regularly $50). Get weekly group sessions & workshops, self-paced courses for all levels, access to tools & templates, challenges with peer feedback, and 24/7 support community. → Join Now
✨ Free 30-Minute AI Consultation - Discover how My AI Fluency can help your business unlock the potential of AI. We'll discuss your goals, explore practical AI opportunities for your industry, and outline clear next steps. → Schedule Free Call
✨ How AI-Fluent Are You? - Test your AI fluency with our interactive quiz. See how you stack up and discover what to learn next. → Take the Quiz
💬 Community | 📞 Book a Consultation | 🌐 Website
Fluently yours, The My AI Fluency Team