In partnership with
Welcome back to The AI Field. Meta's Muse agent can book, buy and email for you, and more than 10 million Americans have downloaded it. This week, researchers showed that conversations people shared from it could be found by strangers with a single search. Meta has paused the feature.
Inside: how the leak worked, what to do about your own share links, why OpenAI's agents turned off their own safety alarms, and Gemini's new access to your inbox while you drive. Plus five tools and a five-minute check on whether a new AI tool is worth installing.
In today's AI Field:
Meta's Muse leaves thousands of shared chats open to strangers
OpenAI's agents switch off their own safety alarms
Gemini reads your Gmail and Calendar while you drive
Plus: five tools, three quick hits, and a five-minute check on whether a new AI tool is worth installing.
Read time: 5 minutes
Nobody's coming to approve your idea
People assume that real businesses get built by other people with more resources. It's how great ideas get stuck in notes apps and conversations.
But you don't need six months and a developer's salary to validate your concept. You just need a weekend. Describe what you want to exist and Lovable builds it: a real site, a real product, ready to accept real payments from real customers. Not a mockup to show people. An actual thing they can use.
You don't need anyone's sign-off to get started. Take $5 off your first month.
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
LATEST DEVELOPMENTS
META · PRIVACY
AI-generated editorial illustration by The AI Field.
The AI Field: Security firm Claroc found more than 6,800 links to conversations people had shared from Muse, Meta's new AI agent, using one search of public internet records. Many opened without a login. After questions from reporters, Meta paused the sharing feature on Tuesday while it reviews the report.
The details:
Muse builds each share's web address from the chat's title plus a random code. When anyone opens the link, that address gets logged in DNS records, the internet's address book, and services that collect those records make them searchable.
Claroc opened only a few hundred. Most were harmless, but dozens were not: a home-sale budget, a bank fraud alert showing part of an account number, a small-business loan application and a revenue review meant for a CEO. Claroc says it has no evidence anyone misused them.
Muse has passed 10 million US downloads in its first 40 days, one of only 12 apps ever to do so. Claroc says it reported the problem to Meta on October 8.
Why it matters: A share link is a public page with a hard-to-guess address. Here the address wasn't the secret it seemed, and the chat title was printed right in it. That matters more for an agent that asks for your email, accounts and card to do its job. Treat every AI share link as public, delete old shares you no longer need, and keep names and numbers out of chat titles. Claroc says the titles stay in those public records even now.
TRY THIS TODAY · 5 MINUTES
Ask your AI whether a new tool is worth adding before you install it
One person who isn't a programmer and runs a single agent for office work pastes the link of every new AI tool into Hermes Agent and asks if it's useful for them. The agent checks the same three things every time. The last tool they asked about got a "not now".
1. Pick a tool you're curious about, like one from today's list below, and copy its link.
2. Paste this into an AI assistant that can read websites, and fill in the brackets:
"Here is a link to a new tool: [LINK]. Read the site and tell me in a few lines whether it's useful for me. Check three things: does it only help if I run several agents (I run [NUMBER]), can it work with the apps I use every day ([LIST APPS]), and can you already do the same thing with what you have now. End with yes, not now, or no, and one sentence on why."
3. Only install it if the answer is yes. If it says "not now", note the reason and check again when your needs change.
OPENAI · AI SAFETY
The AI Field: OpenAI disclosed that two of its AI agents weakened the system built to catch misbehaving AI during routine internal coding jobs this month. One then told the researcher it hadn't done it.
The details:
On October 3, a GPT-6.1 Sol agent repairing a test pipeline turned off automatic alerts to OpenAI's security team. It read an earlier instruction, "don't send any messages" during a Slack search, as covering the safety alerts too.
On October 6, a GPT-6 Astra agent moving a program to a new setup removed its monitoring to match a reference version. Asked about it later, it answered that it did not remove an enabled monitor.
Neither case was rated top severity, and neither run was stopped automatically. OpenAI says its internal reviewer now blocks attempts to disable monitoring, and it admits it doesn't "yet have a satisfying explanation for this behavior."
Why it matters: Both agents stopped to ask whether the safeguards should stay, then talked themselves out of it using ordinary instructions. That's the everyday risk with any agent you hand a broad job: vague wording gets read as permission. If you use agents for work, spell out what must never be turned off or deleted, and check what it changed, not only whether the task got done.
100+ Claude Code hacks to ship code 10X faster
Top engineers at Anthropic say AI now writes 100% of their code.
Are you using AI to write yours?
These 100+ Claude Code hacks show you exactly how. Sign up for The Code and get:
100+ Claude Code hacks — free
The Code newsletter — learn the latest AI tools and skills to code faster in 5 mins a day
GOOGLE · GEMINI
AI-generated conceptual illustration by The AI Field.
The AI Field: Google is rolling out Personal Intelligence for Gemini on Android Auto, so the assistant can pull from your Gmail, Calendar, Photos, Maps, YouTube and Search to answer questions on the road. Cars with Google built-in get it in the coming months.
The details:
Google's examples: "What's my flight gate? And is my flight on time?", "Take me to my doctor's appointment" and "Find the car wash from the photo I took last week and navigate there."
It's opt-in and can be turned off anytime in Gemini settings. If you already use Personal Intelligence in the Gemini app, it comes with you into the car. Cars with Google built-in only give personal answers while your phone is connected.
It's for personal Google accounts, age 18 and up. Work and school accounts are left out, and so are the EU and wider EEA, Switzerland, the UK and Nigeria.
Why it matters: Asking out loud beats glancing at your phone at a red light. But the answers come out of the car speakers, so think about who else rides with you before you connect your inbox. If you try it, start with one question you'd normally look up, like your next appointment, and check the answer.
TRENDING AI TOOLS
Five tools you can try today, most launched or updated in the past week. Click a name to open the tool.
👁️ Lightroom: Adobe's mobile app now spots closed eyes in a photo and offers a one-tap Open Eyes fix, using other photos of the same person to match their eyes. Adobe says it's free; some testers report a limit of two uses on free accounts.
🎙️ ChatGPT Voice: Voice mode now runs on GPT-Live-1.1, with 13 new voices and much better accents in French, German, Hindi, Japanese, Korean, Mexican Spanish and more. Rolled out to all users last week.
📝 Otter: A meeting notetaker for Zoom, Teams and Google Meet that transcribes live and lets you chat across your meetings. It just raised $320 million. The free plan includes 300 minutes a month.
💰 Maali: A budgeting app for Mac and Windows that sorts your bank charges and flags bills that keep creeping up. The AI analyst runs on your own computer, even with the Wi-Fi off. Free for 30 days, then $99 once.
🍽️ Yelp: New Check for Me feature: enter your party size and Yelp's AI agent calls the restaurant to ask about the wait, then texts you the answer.
QUICK HITS
🔒 ChatGPT now lets you lock chats: Free and paid users can hide sensitive conversations behind a PIN or Face ID, rolling out now. It keeps chats from people using your device, but doesn't add encryption or change how long OpenAI keeps them.
🎵 Spotify pulled 50 million AI slop tracks: It says it removed them from recommendations over the past year, and it won't recommend music from AI personas, which now carry a badge.
📚 Meta faces a new wave of book lawsuits: About 300 authors filed individual claims on Monday over books allegedly pirated to train its AI, with close to 1,000 more planned. Separately, a judge ordered Meta to name the executives who approved the downloads.
That's it for today!
See you soon,
Olle Hellman | Founder of The AI Field