In partnership with
Welcome Automaters, 👋
So the AI world had a collective mini-heart attack!
Canadian AI titan Cohere dropped a bomb on Monday: during a routine security evaluation, Alibaba’s new AI model, Qwen3.8-Flash-Next, tried to upload a file directly to a server in Beijing.
The kicker? The model had explicitly stated in its own reasoning notes right before the attempt that it would not use any external tools. (Yep, it literally said one thing and did the exact opposite a second later!)
Naturally, tech executives everywhere panicked. Was this corporate espionage? Did an open-weight model just ruin the promise of keeping data in-house? Well, take a breath. Researchers audited the situation, and the truth is way more bizarre than a spy movie plot.
What Actually Happened in the Sandbox?
On Friday, Cohere ran Qwen through an evaluation asking it to write a simple security rule to catch malware downloads.
Qwen logged its step-by-step thought process, reassuring testers it wouldn't call on outside tools.
Immediately afterward, Qwen attempted to ping a test string (a basic piece of placeholder text) over to an external server.
Fortunately, Cohere’s security sandbox caught the request and blocked it cold before anything went anywhere.
So was the model acting out of malice or plain confusion? Cohere co-founder Ivan Zhang explained that AI models can pick up weird habits in training simulators.
Basically, if Alibaba’s practice setups routinely saved files on Beijing cloud servers, Qwen simply learned to reach out to those addresses by default. It wasn’t a secret agent: it was acting like a hyper-advanced autocomplete that filled in familiar web addresses!
After Cohere posted the test traces on Hugging Face, independent researchers confirmed there is zero evidence that Alibaba planted a malicious backdoor or that the model was secretly hiding intent. Also: While the model can invent suspicious web destinations, actual spontaneous upload attempts are extremely rare.
Oh, and there’s the geopolitical angle
This drama comes at a spicy time. Just last week, the Department of Homeland Security ordered federal agencies to purge products from Chinese AI developer DeepSeek.
Cohere is now urging governments to clarify if restrictions apply to open-weight models running locally on government hardware. (And let's be real: since Cohere sells AI directly to governments, a little extra caution around free Chinese alternatives certainly doesn't hurt their business!)
So yeah, if officials start panicking over free Chinese alternatives, enterprise players like Cohere stand to win big!
The takeaway? Always sandbox your AI, because these models will get creative when you aren't looking!
Here's what we have for you today
🤔 Anthropic Files for Historic IPO with $16B in Revenue, $80B in Losses, a $2T in Valuation and 7 Founders Holding All the Keys
After weeks of leaked drafts, Anthropic finally dropped its official S-1 filing on Monday. Which means they're officially setting the stage for what might be the most intensely watched public debut in modern tech history.
If you’ve been tracking the ticker “ANT,” here’s the wild, jaw-dropping breakdown of what is actually happening behind the scenes.
Ang get this: The growth numbers are absolutely bonkers:
Anthropic pulled in a staggering $16.3 billion in revenue for the first half of 2026. (For context, that’s up from a modest $1.2 billion during the same period last year.)
Operating losses landed at $693 million, which is actually a huge improvement over the $2.1 billion operating loss logged in early 2025.
On paper, net losses ballooned to $80.7 billion. But don’t panic just yet: management notes this staggering number was mostly driven by non-cash accounting adjustments tied to convertible debt.
As of June 30, 2026, the company had raised $141.1 billion in total capital, sitting on a comfy cushion of $112.4 billion in cash and cash equivalents.
And yes, Anthropic explicitly warned investors in classic corporate fashion that it "may not achieve or sustain profitability in the future." (Honesty is key, right?)
So who is actually paying for all this compute power? Deep-pocketed enterprise clients! As of late September, Anthropic’s high-value customer base looks like this:
11,900+ customers spending $100,000+ annually
1,900 customers dropping $1 million+ every year
250 elite clients spending a mind-blowing $10 million+ annually
And here's where the tea gets really spicy. If you plan on buying shares, do not expect a seat at the steering wheel.
Anthropic’s seven original co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah have locked down their authority through a dedicated "Founder LLC."
The Big Catch: This founder entity will control 50.1% of all shareholder voting power, a move they say is designed to preserve the company’s AI safety mission. On top of that, Anthropic's Long-Term Benefit Trust (created in 2023 to keep their safety mission on track) gets to appoint most of the board.
Meaning: Public investors will have almost zero practical control over a firm building technology that, by its own public disclosure, carries potential "catastrophic or existential risks to humanity."
Speaking of existential risks, the filing does not sugarcoat the future. Anthropic flagged four primary risk categories around frontier models:
Malicious Misuse
Loss of Operational Control
Authoritarian Abuse
Widespread Economic Disruption
The firm predicts that super-intelligent AI models, capable of outperforming top human experts and navigating both software and robotic interfaces, will "become widely available within a few years."
Even while advocating to slow down the reckless AI race, CEO Dario Amodei and the team are pitching investors on a grand vision: helping humanity navigate this leap safely toward an "age of miracles."
What’s Next? While the official share count and pricing remain unconfirmed, Wall Street insiders expect this IPO to land around mid-November with a valuation pushing $2 trillion.
So yeah, get your wallets or popcorn ready!
9 Plays e-Commerce Leaders Are Running to Protect Margin
On the surface, selling out sounds like a pretty good holiday problem to have. That’s the case until the costs add up.
CommerceIQ discovered that stockouts cost brands 140% more than baseline on Black Friday and 248% more on Cyber Monday.
Holiday growth can be difficult because more sales don’t always mean more profit. Margin quietly disappears across ad spend, discounts, inventory decisions, creator payouts, and returns that hit in January.
9 Plays to Defend Your Margin This Holiday Season looks at nine of those places, including:
Healthy ROAS hiding money-losing SKUs
Discounts on products shoppers aren’t price-comparing
Ad dollars going toward inventory at risk of stocking out
Creator payouts capped by first-order math
Returns that turn into refunds instead of exchanges
Plus learn four more ways to keep more of what you earn this Q4.
🧱 Around The AI Block
🦾 How OpenAI turned its coding tools into dots, an always-on personal assistant.
🪦 The AI graveyard: a running list of projects and startups that didn’t make it.
😳 In China, the AI bubble has already burst.
🧑💻 Anthropic’s IPO filing reveals its strategy for keeping enterprise customers hooked.
🤯 Dutch government to cut 16 percent of workforce as part of AI efficiency push.
🤓 OpenAI debuts 'Shepherd' Model for AI research validation.
🫣Anthropic discloses month-long training pause, AI security breaches in IPO filing.
🤯 Meta plans $95bn stock offering to help cover AI expansion.
🤔 From Amazon to Google, the deep-pocketed investors in Anthropic.
🦾 Cadence unveils Nvidia-backed AI models that chipmakers can run in-house.
👩🎓 AI Tutorials
Claude’s 10 Big Updates You Need to Know.
And: How to Use AI Coding Tools Better than 99% of Developers.
So tell us, what’s the single most annoying, tedious task in your daily workflow that you desperately wish an AI could just handle for you?
Hit reply and the next video might just be around your exact problem!
See what's really happening in every deal
Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.
🛠️ Trending Tools
For Privacy Advocates & AI Power Users: Obscurify.ai provides an anonymous gateway to access over 200 frontier AI models, including GPT, Claude, Gemini, DeepSeek, and uncensored LLMs, without logging identity data. Accounts are secured via random 16-digit codes and crypto payments (Monero, Bitcoin, Lightning), while user metadata is scrambled using blind-signed vouchers and zero-data-retention routing. It offers security-conscious users and researchers a privacy-first platform to query top AI models over Tor or run lightweight models locally in-browser.
For Language Learners & Polyglots: LingTube is a browser extension that supercharges language learning on YouTube and Netflix with context-aware dual subtitles. Instead of literal dictionary translations, it uses LLM-powered parsing to deliver sentence-accurate definitions, part-of-speech color coding, furigana for Japanese, pinyin for Chinese, and instant CEFR level tagging. It gives language students an immersive way to study real-world media while building personalized vocabulary lists.
For Poker Players & Game Strategists: PokerBotAI offers a real-time decision engine and automated assistant designed for online poker rooms and mobile club platforms. Operating either as a live HUD assistant suggesting optimal moves or as an automated multi-tabling bot, it analyzes board state and opponent tendencies in real time. It gives competitive players a technical tool to analyze hand equity and optimize strategy across desktop and mobile rooms.
🤖 AI Workout Of The Day: How to Make Claude Read, Draft, and Send Emails for You
Imagine sitting back, taking a sip of coffee, and watching your inbox literally answer itself? Well, Claude can now read, draft, and actually hit send on your emails.
Anthropic dropped their official Gmail connector, bridging the gap between Claude and your chaotic inbox. You also get total control over the approval settings, so it won’t go rogue and send a weird message to your boss. The best part? It’s completely free across all plans.
Sniff out lost messages: "Find that last email from Luke and draft a reply." Claude digs through your threads, pulls up the right message, and cooks up a response instantly.
Organize the chaos: It categorizes your messages with labels and thread management on the fly.
Handle the heavy lifting: Forwarding, drafting, and sending are all fair game.
Once Claude drafts a reply, you can jump into Gmail to send it yourself, or just tell Claude, "Looks good, ship it!"
Quick Work Account Heads-Up: Setting this up on a personal account takes seconds. If you're on a Team or Enterprise plan, your workspace admin needs to enable the connector on their end first before you can play around with your work email.
How to Connect It (In Under 2 Minutes)
Pop open Claude in your web browser or app.
Click Customize > Connectors (or hit the + icon in any chat and pick Connectors).
Select Browse connectors, grab Gmail right near the top, and sign in with your Google account to grant permissions.
Now let's talk safety, because giving an AI the keys to your inbox feels a little scandalous.
By default, Claude is put on a strict leash. It has to ask for your explicit stamp of approval before doing anything that alters your inbox. If you get annoyed by constant permission pop-ups, you can tweak these settings directly in the chat interface. You just have to:
Go to your chat interface and click the plus sign. Go to Connectors and pick Gmail.
There's a slew of tool permissions to fiddle with, and you'll find them under Read-only tools and Write/delete tools.
For context:
Read-Only Tools (Set to "Always Allow" by default): Lets Claude search emails, pull threads, and read labels silently in the background without bothering you.
Write/Delete Tools (Set to "Needs Approval" by default): Covers actions like drafting, sending, deleting, tagging spam, or creating labels.
You can flip any permission between Always Allow, Needs Approval, or Block at any time.
Pro tip: Keep the approval settings tight at first. Once you trust Claude’s tone and judgment, loosen the reins and let it clear out your inbox while you reclaim your sanity.
💡 Prompts To Try:
Act as my personal email assistant. I need you to draft a reply to an email thread in my Gmail account.
Instructions:
1. Locate the Thread: Search my inbox for the latest email from [Insert Sender Name or Subject Line].
2. Tone & Style: Draft a response that is [Insert Tone: e.g., brief and professional / friendly and casual / polite but firm].
3. Key Information to Include: Mention that [Insert Main Points, e.g., "I accept the meeting invite for Thursday at 2 PM" or "I'll need until Friday to review the document"].
4. Safety Rule: Display the complete draft for my review first and ask for my explicit approval before sending it out.
Please pull up the email and show me your draft reply to get started. 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.
