In partnership with
Welcome back to The AI Field. For two weeks, a security researcher watched one hacker break into more than 100 company websites, most of them online shops. The hacker did almost none of the work. AI agents did the searching, the trial and error and the break-ins, for roughly $50 each.
Inside: what that means for the shops that have your card, why the US government just pulled DeepSeek off its computers, and the race for money between OpenAI and Anthropic. Plus seven tools and a five-minute language lesson.
In today's AI Field:
A hacker sends an army of AI agents into 100+ company websites
The US government pulls DeepSeek off its computers
OpenAI races to raise $30 billion before Anthropic's IPO
Plus: seven tools, three quick hits, and a five-minute language lesson.
Read time: 5 minutes
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
LATEST DEVELOPMENTS
AI AGENTS · CYBERSECURITY
AI-generated illustration by The AI Field.
The AI Field: Security firm NorForge found more than 20 unprotected control panels for Cairn, an open source tool that coordinates AI agents, being used for break-ins. Through them, researcher Snorre Fagerland watched a Chinese-speaking hacker hand the agents a list of targets and let them work. Victims identified from the hacker's own dashboard include Deere, SharkNinja and Cummins.
The details:
From September 25, the agents tried more than 1,200 online shops in the US, Canada, Australia and Japan and got into over 100. The instructions were under 100 Chinese characters. The goal was hidden, lasting access to store and payment systems, to cash in on later.
A successful break-in took just under 23 hours on average. Fagerland estimates the campaign burned about $5,000 worth of AI, or roughly $50 per hack, apparently running Claude Code with DeepSeek's low-cost model.
Deere says it saw no sign that sensitive data was accessed, and SharkNinja says only a product support site was hit. Another hacker spotted the same way stole a trove of credit card data.
Why it matters: Breaking into a website used to take skill and patience. Now one person can rent both by the hour. Only about one in ten attempts fully worked, by NorForge's estimate, but at a few dollars a try that hardly matters. You can't fix a shop's servers. You can limit what a breach costs you: unique passwords, passkeys where offered, and fewer saved cards.
TRY THIS TODAY · 5 MINUTES
Practise a new language with lessons built from your own week
One person set up a Portuguese tutor in OpenClaw that writes exercises around their real life. Their padel plans turned into a grammar lesson, and a pile of paperwork turned into pronoun practice.
1. Paste this into your AI assistant and fill in the brackets:
"Be my [LANGUAGE] tutor. My level is about [LEVEL]. Each morning, look at my calendar and notes for the day and write a 10 minute lesson around what I'm actually doing: 5 phrases I might need, one grammar point that fits them and 3 short exercises. Correct my answers when I reply, and keep a list of mistakes I repeat so you can bring them back in later lessons."
2. If your assistant can't see your calendar, add a few lines about what you're doing today under the prompt.
3. Answer the first exercise and send it back. It corrects you and keeps track of the mistakes you repeat.
DEEPSEEK · US GOVERNMENT
DHS headquarters, September 2021. Photo: Glenn Fawcett / CBP.
The AI Field: The Department of Homeland Security has ordered federal civilian agencies to remove everything made by Chinese AI company DeepSeek and to stop buying it. The order went public on Monday, but it has quietly been in force since late July.
The details:
It covers outside apps and services built on DeepSeek's models too, not just DeepSeek's own app. Agencies like Treasury, State and Justice are included. The Defense Department and intelligence agencies are not.
The signed order points to data kept on servers in China, where the law lets the government demand it, and to the prompts and answers users send. DeepSeek's own privacy policy says it collects, processes and stores personal data in China.
DeepSeek is one of the most downloaded AI apps, and its cheap models also run inside services from US companies like Snowflake, Vercel and OpenRouter. Agencies that want an exception have to ask 60 days ahead.
Why it matters: The ban doesn't apply to you at home, but its reasoning does. Anything you type into DeepSeek's app or website is stored in China, so keep work files, health details and ID numbers out of it. The same low-cost model turned up in today's lead story, a reminder of how far cheap AI travels once it's out.
When your support agent gets it wrong, who's accountable?
Every agent handling refunds, account changes, or tickets needs three answers: what it can pick up, what it's allowed to do, and when it hands back. Running Agents in Customer Work is four conversations on agentic AI in customer ops. Register now, four Tuesdays, 10 a.m. PT.
OPENAI · ANTHROPIC
AI-generated illustration by The AI Field.
The AI Field: OpenAI is reportedly rushing to raise about $30 billion at a value above $1.4 trillion, and wants the money in hand before Anthropic's stock market debut, expected in mid-November. Existing backers SoftBank and Abu Dhabi's MGX are said to be in talks to join.
The details:
Both companies are chasing the same big investors. People familiar with the plan say OpenAI wants to close before Anthropic's public filings invite side-by-side comparisons.
OpenAI is telling investors it now brings in $50 billion a year at its current pace and expects $70 billion by year-end. It was valued at $852 billion in March.
Anthropic's IPO price may come down: some investors are pushing for about $1.5 trillion instead of $2 trillion after its third quarter fell short of hopes. It is also reportedly close to another $45 billion computing deal with Nscale.
Why it matters: The fight for money is also a fight over prices. OpenAI says cutting prices on its cheaper models boosted use and that it can cut further, which is good news if you pay for AI. The risk sits elsewhere: chip stocks lost $1.6 trillion in value in under two weeks as investors questioned the cost of the build-out. If you're tempted to buy Anthropic shares on day one, read the prospectus first and decide your price before the hype starts.
TRENDING AI TOOLS
Seven tools you can try today, most launched or updated in the past week. Click a name to open the tool.
🖼️ Krea 3: Krea's new image model is built for edits: restyle a product shot, swap a background or keep the same character across images, in seconds. The free plan includes about five Krea 3 images a day.
⚙️ Make: A no-code automation platform that connects 3,500+ apps and AI models, so you can build workflows and AI agents by prompt or drag-and-drop. The free plan includes 1,000 credits a month, no card required.
🧵 Wabi: A personal AI assistant from Replika's founder that you can share. Each task gets its own thread with small custom apps, and you can invite a partner or family into a thread. Free for now on iPhone, Android and Mac.
🗒️ Gemini Notebook: Google's research notebook, formerly NotebookLM, now shows you its plan before it turns your sources into slides or a video overview, so you can reorder or add sections first. Free, rolling out this week.
🔎 Quello: Connect your docs and tools, then ask questions by chat or voice. Every answer links to the source it came from, and if your files don't say it, neither does Quello. Free to start.
📋 readywhen: An assistant for people who run their own business. It catches what you promise in emails, calls and messages, reminds you who's waiting and drafts the follow-ups. Free to start.
👽 Tolan: An AI companion with a friendly alien character that helps you talk through hard days and get your to-dos done. Free to download on iPhone; full features need a membership.
QUICK HITS
🛡️ Microsoft says AI models can't be trusted: "Frontier models, regardless of their origin, cannot be trusted," its security chief wrote, launching Guardian Agents that watch a company's AI agents and can cut their access, ask a human to approve or shut them down.
💬 WhatsApp's help pages hid rules for its chatbot: About 80 pages exposed hidden instructions for Meta's support bot, including one telling it not to suggest contacting a data protection regulator. Meta says a bug caused it and the text is gone.
📰 Canada's biggest newspapers sue Anthropic: Postmedia, Torstar, The Globe and Mail and The Canadian Press say Claude was trained on their journalism without a license, partly by getting around paywalls. The claims are unproven.
That's it for today!
See you soon,
Olle Hellman | Founder of The AI Field