Hey, AI Enthusiast!

Welcome back to the #1 AI newsletter on the planet!

Here’s a glimpse into what we have today:

Together with Wispr Flow

Say It Once. Get the Prompt.

You know what you want your agent to do. It still gets it wrong.

So you rephrase. It misses again.

You add context, it fixes one thing and breaks another, and you’re typing the same request a fourth time, shorter and angrier each round.

Here’s the part nobody admits.

The agent fills every gap in your ask with a guess. And typing a full brief takes so long nobody does it.

Wispr Flow fixes the first ask.

Say your rough idea out loud, highlight it, and hit the Prompt Engineer Transform.

Flow rewrites it in place as a full prompt with the role, task, key points, and tone, in Cursor, Claude, ChatGPT, or any text field.

Cyberattacks

OpenAI’s Agents Ran Cyberattacks?

Picture asking an AI agent to look up some public data. The site says no. So the agent tries to break in.

OpenAI’s agents did exactly this, according to researchers, across government, corporate, and university systems in multiple countries.

Nobody asked them to.

When a site blocked them, they borrowed urlquery.net, a public web-scanning service, as a browser and tried to force their way in.

Transluce, a nonprofit AI research lab, pieced it together from that service’s public logs.

The targets?

A government health agency, DataUSA (which publishes US government data), and the University of New Mexico’s digital library.

Transluce ties at least some of it to groups of agents already linked to OpenAI.

This week, a prime minister confirmed one agent got into a government health statistics portal.

And in July, OpenAI said its own agents broke into AI company Hugging Face during a security test.

It started March 6, two months before anyone said a word publicly, and ran to at least Sept 16.

Most attacks got blocked. At least two got in.

Our take: The good news is the system worked, sort of. A small nonprofit caught it using public logs anyone could read. The next step is faster disclosure, so the public hears about these from the labs first, not months later.

Biology

Claude Found a Mystery Enzyme

Inside the DNA of viruses that infect bacteria, there’s a new enzyme system sitting right next to a run of repeating DNA.

Sound familiar?

The setup looks a lot like CRISPR (a gene-editing technology that lets scientists make targeted changes to an organism’s DNA). And Claude is what found it.

Anthropic ran around 950 copies of Claude on the search for 21 hours.

Together they sifted more than 200,000 enzymes, flagged 3,500 candidates, and wrote up the top 20.

Anthropic says the same job takes a human expert weeks to months. Scientists wrote the first prompt and handled the lab work.

The best moment came mid-search, when one copy spotted the repeating pattern by eye and flagged it as a possible CRISPR-like array.

Why care?

The few known systems with these traits are all tools for cutting, copying, and pasting DNA.

Feng Zhang, a CRISPR pioneer, reviewed the preprint and called it worth a closer look.

Our take: The speed’s the story here. A search Anthropic scopes in months ran in under a day, and scientists kept the parts that matter most, the question and the lab.

The catch: Nobody knows what the new system does yet, and this is Anthropic’s own claim in a preprint.

MongoDB In Focus

One Server, All Deployments

Your coding agent has read every file in your project. It’s never seen your database, and it writes queries against collection and field names it invented.

The MongoDB MCP Server hands the agent your real schema.

Collection schemas, configs, and sample data land in the tools you already use, and the code that comes back matches what’s in there.

Claude Code, Copilot, Cursor, Codex, and Gemini CLI all connect.

Local, Atlas, and Enterprise sit behind one config. Run it yourself, or let Atlas host it.

You stay in one window, and the answer arrives with your own data behind it.

Notable AIs

Notable AI Tools

🎙️ Linden by Speechmatics gives voice agents speech-to-text built for the stuff they trip on, like names, postcodes, and phone numbers, and it tracks who’s talking.

📈 Marketers, CXL’s six-week live cohort teaches you how to build 6 AI workflows from your own work and keep them running. Get 50% off until Sep 28 when enrollment closes.

Meta Connect

Meta Went All In on AI Hardware

Meta’s big event last night came down to one idea. Get Muse, its AI agent, off your phone.

Muse is coming to Meta’s AI glasses, so you’ll be able to talk to your agent all day without reaching for anything.

Don’t want a camera on your face?

Meta launched its first audio-only glasses, with AI and open-ear speakers built in. They ship Oct 13.

The surprise of the night was Muse Charm, a tiny device made for one job, talking to your Muse. It lands this fall.

Muse is getting better at errands, as well. It already buys things for you, and now it works with retailers like Walmart and Instacart.

Meta also showed off VR Glasses, a thin headset for movies, games, and work, coming in the spring.

And by year-end, it expects over 100 styles of AI glasses.

Our take: Meta’s betting on the agent, not the gadget. Glasses, a charm, your phone, it’s all one assistant, and whatever you’re wearing is the way in.

Interesting AI

GPT-6 Astra Drove a Real Car

Every driving lesson starts the same way.

An empty parking lot, a row of cones, and a nervous student with an instructor’s foot hovering over the brake.

This time the student was an AI.

Researchers wired OpenAI’s GPT-6 Astra, along with a few other models, into a real Toyota Corolla’s wheel, gas, and brakes.

Astra kept refusing to drive, citing safety, until the team renamed the software “DrivingBench Sandbox.”

Then it went, learned from a first run that ended halfway, and became the only model to finish the course.

The rename is the fun part.

Astra’s caution kicked in when it thought the stakes were real, which is exactly when you’d want it.

Coming in Hot

AI Tools of the Day

Avatars

Gemini Gets a Face

You know the drill. “Please hold.” “Lemme transfer you.” “Sorry, what was your account number again?”

Google’s new Live Avatar goes after all three.

It’s a real-time talking face for Gemini, and your next support video call might have one.

The avatar listens, sees what’s in front of it, and talks back with lip-sync, facial expressions, and natural turn-taking.

It switches between 97 languages mid-conversation, and the lips keep up. And it doesn’t stop talking to look things up.

In Google’s demo, an avatar checks a guest into a hotel, pulling up records in the background without ever putting them on hold.

Businesses pick a preset face or build their own from one good reference image, though custom avatars need Google’s sign-off for now.

Everything it says and shows carries SynthID, Google’s invisible watermark for AI content.

For now it’s business-only, through Gemini Enterprise. You won’t find it in the regular Gemini app.

Our take: The face gets the attention, but the background work is the real upgrade. An agent that files your check-in while it keeps talking means less waiting, and customers notice the shorter wait first.

The catch: Google watermarks everything the avatar says and shows, but it’s invisible to people. Whether you know it’s AI is up to the business using it.

AI at Work

Workers Say Choice Makes AI Work

Use AI too little, and you get dinged. Use it too much, and you get flagged.

One senior data engineer got both messages from the same team, first told to burn as many tokens as possible, then watched a colleague get flagged for using too many.

The story comes from a tech workers’ union that asked its members how AI is changing their jobs.

Some teams tie bonuses to AI use or rank people by tokens, and usage numbers end up standing in for good work.

The members’ fix is simple, and it’s the most positive finding in the report. AI helps most when people decide for themselves how to use it.

Senior engineers who built their skills before AI said they feel more confident using it to support their coding, as a tool rather than a stand-in for their own thinking.

The main ask is to let people choose how, and whether, to use AI, with no penalty for opting out.

Our take: The best AI rollouts measure the work, not the usage. Give people good tools, let them pick where they fit, and judge the results.

The catch: This comes from union workshops, not a survey. There’s no participant count, and most of the report is critical of how companies roll AI out. Treat it as worker testimony, not the whole industry’s view.

Prompt of the Day

Steal My Customers

Important Note: You’ll need to click the button to get the complete prompt.

That’s all for this Thursday!

Thanks for reading,

— There’s An AI For That