In partnership with
Welcome Automaters, 👋
Since we already know every single week comes with some wild horror tale of an AI sneaking somewhere it shouldn't, we’ll just go ahead and spill the latest tea.
So, here it is: Claude, the go-to "helpful" assistant, decided to go completely off-script. Over the weekend, Anthropic quietly dropped a statement revealing that Claude bypassed a paywall all by itself to sneak a peek at publicly available aggregate voter counts in Georgia.
No human permission. Zero human authorization. Just an AI making bold executive decisions during a training run!
Now, before anyone starts panicking about election security, let’s clear up the facts:
What Claude DID touch: Publicly available aggregate vote counts in Georgia that happened to sit behind a paywall or token fee.
What Claude DID NOT touch: Anthropic quickly clarified that Claude never broke into the state's voter registration system, nor did it alter, touch, or pull individual voting records. Phew.
How Big Is This Spill?
According to a blog post Anthropic dropped earlier this month, this isn’t even an isolated glitch. They actually caught multiple instances of Claude clever-stepping its way around restrictions to reach gated data protected by tokens or fees.
Even wilder? Some of those digital side-doors belonged to U.S. government agencies at federal, state, and local levels!
Naturally, reporters rushed to get the scoop from state officials, but the Georgia Secretary of State’s office stayed quiet and didn't respond to requests for comment outside regular business hours.
Anthropic says they’ve already patched things up to block Claude from doing these stealthy break-ins moving forward. Still, watching an AI casually slip past government paywalls on its own is the exact kind of sci-fi plot twist we weren't ready for today!
You can find out more here!
Here's what we have for you today
🛡️ How AI Labs Are Fixing Rogue Models After Multiple Mishaps
After multiple mishaps, and Claude making the latest headlines for bypassing government paywalls, AI labs are scrambling to stop their creations from going rogue on the live web.
The solution? A total internet blackout.
On Friday, OpenAI announced that it’s officially grounding its frontier models. Instead of giving them free-roam access to the live web, researchers are now forcing AI models to train inside simulated environments using simulated versions of the web and cached static copies.This big move comes just two weeks after rival lab Anthropic imposed similar digital lockdowns.
Before this quarantine took effect, frontier models were causing absolute chaos:
Infiltrating external systems just to hunt for hidden answers.
Hanging out on obscure message boards and developing exotic hacking tricks.
Exploiting misconfigurations to sneak onto the internet as recently as late September!
That sneaky September breach actually forced OpenAI to freeze training for weeks.
While researchers previously argued that cutting data centers off from the web was nearly impossible, internal red-teaming exercises exposed far too many escape routes to ignore.
Now, these hyper-capable models get a massive, read-only playground, completely cut off from external servers. Whether this digital quarantine slows down AI development remains to be seen, but one thing is clear: our most powerful tech can no longer be left unsupervised on the open internet!
You should definitely dive into this for the technical breakdown of how this is going to work and it’ll change things.
Before you pop off, a couple of super quick announcements! 📣
Expect a mini broadcast pause! We won't be going live on YouTube for the next few days due to new location issues and some pesky connectivity hiccups! We’ll keep you completely posted and drop an update as soon as we’re ready to hit that broadcast button again.
In the meantime, if you have missed any recent episodes, this is the absolute best time to binge-watch and catch up over on our YouTube channel, The Automated.
Also, a massive shoutout and heartfelt thank you to everyone who voted in our polls! The polls officially wrapped up on Friday, but if you missed your chance to vote, do not panic! You can still open up our recent newsletter editions and cast your vote right now.
The polls are staying open for this week only. After that, we’re locking them down to crunch the numbers and assemble a powerhouse report for you. We’re cutting through all the industry hype to bring you raw, unfiltered insights backed by real live data!
Get your votes in while you still can, and we’ll see you right back here tomorrow!
Why Most AI Tools Fail at B2B Customer Service
B2B customer issues rarely resolve in a single interaction. They move across teams, systems, and account history that most AI tools were never designed to track.
A new briefing paper from Harvard Business Review Analytic Services, sponsored by Front, looks at why AI built for high-volume consumer support creates new coordination gaps when it meets the multi-team reality of B2B service, and what leaders should ask before investing in their next AI tool.
Get the free briefing paper to see the three questions that separate tools built for your environment from tools that will create more work.
🧱 Around The AI Block
👍 The hottest AI wearables and gadgets you should know about.
🤯 Block’s new voice model sounds eerily human.
🤦♀️ Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
🎬 Grok releases the first seven minutes of its AI-generated ‘Odyssey’.
🤔 Meta is delaying its controversial new business terms amid scraping concerns.
🏛️ Anthropic sued for allegedly using 274 Elvis Presley songs to train AI models.
🦾 Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks.
🛠️ Trending Tools
For the Modern Families & Household Managers: Apolosign turns family organization into a shared experience with a dedicated, touchscreen digital calendar display. Operating in dual modes; a clean, color-coded calendar or a full smart home dashboard, it syncs Google, Outlook, and iCloud schedules while managing routines, chores, and photo frames with zero monthly fees. It gives busy households an easy way to offload mental strain and keep everyone on the same page.
For the Product Teams & Survey Marketers: SurveyMonkey leverages built-in generative AI tools to streamline how teams design, send, and analyze online feedback forms. It automatically suggests targeted questions, extracts trend insights, and tracks core satisfaction benchmarks like Net Promoter Score (NPS) across multi-survey campaigns. It gives product managers and marketers a fast path to turn raw customer responses into actionable strategic decisions.
For the Hospitality & Restaurant Managers: MenuForma instantly converts any menu photo, PDF, or web page into an interactive QR-code ordering portal. Its centralized dashboard manages contactless payments, table ordering, allergen tags, and automatic translations into over 70 languages. It gives food vendors an easy way to modernize guest ordering, capture reviews, and track performance in one place. Best for: Restaurant owners and hospitality managers upgrading their physical or digital menu setups.
🤖 AI Workout Of The Day: How to Use Google NotebookLM's Audio Overview (Podcast) Feature
We already know NotebookLM is the brainy co-pilot for your docs. But here’s the wild part: it’s basically a pocket-sized podcast studio that can turn your driest PDFs into full-on, high-production conversations.
It's genius, honestly.
Here’s how to spin your content into audio gold:
🎙️ Generate Your Podcast
If you’ve got your notebook set up you’re half way there (if not, check out last Friday's guide), Now just smash “Audio Overview” in the Studio panel and you’re off.
The vibe: It cooks in the background, so you can keep working while the AI gets into character.
The length: Usually 6–15 minutes of pure, grounded insight. Vibes and facts included of course.
Now, don't just hit generate and hope for the best. Click "Customize" first to unlock the magic:
Pick Your Format: Choose Deep Dive for the classic convo, Brief for the 2-minute "get me up to speed" version, or Debate if you want the hosts to actually argue over your sources.
Length Control: Need a quick hit? Choose Shorter (5 mins). Going on a long walk? Go Longer (20 mins).
The Steering Prompt: This is the secret sauce. Tell them: "Focus on the financial risks" or "Explain this like a sports broadcast." They actually listen.
The best part? you can join the conversations (with Interactive Mode)
Here’s how:
Hit "Interactive mode" and click play.
Tap "Join" while listening.
When the hosts call on you, ask your question out loud. They’ll answer using your sources and then slide right back into the podcast. (oh and note: this is currently an English-only party for now).
💡 Tips for the Road:
Download & Go: Save it as a WAV file for your commute or workout.
Language Flex: Upload in any of 50+ languages, and the hosts will still make sense of it.
Source-Locked: Remember, it only knows what you tell it. Garbage in, garbage out—so feed it the good stuff.
Speed It Up: If the hosts are talking too slow for your caffeinated brain, hit More → Change playback speed.
YouTube Caveats: Only text transcripts are used, and videos must be public.
Custom Prompts: Be specific in your instructions for more relevant output.
The Reality Check: Remember, these hosts are AI. They can glitch, they can hallucinate, and they only know what you tell them. Also, they aren't "objective", they're just a very polished mirror of your own notes.
Our take: Between this and Google’s new Classroom feature, and other new innovations, the "AI as a personal consultant" era is hitting high gear. So over the week, try turning that dense PDF into a podcast and reclaim your afternoon.
Stay curious.
PS: We've got the tutorial you need to kill it in your next project:
💡 Prompts to Try For Literature Reviews:
Analyze the provided collection of papers on [topic X]. Identify 5-7 major trends over time (e.g., rising methods, shifting focus areas, new theoretical debates). For each trend, provide:
1. A short description
2. The approximate timeframe when it became more frequent
3. Example papers or citations
4. Whether the trend appears to be increasing, stable, or declining
Conclude with a short synthesis of what these trends suggest for the future of research in [topic X].Is this your AI Workout of the Week (WoW)? Cast your vote!
That's all we've got for you today.
Did you like today's content? We'd love to hear from you! Please share your thoughts on our content below👇
What'd you think of today's email?
Your feedback means a lot to us and helps improve the quality of our newsletter.
