The Download

Terminator 2: Judgment Day/TriStar Pictures via Giphy

TL;DR: Yesterday, Anthropic released a report saying it’s stopped several potential attempts to use Claude for bioweapons development (among other nefarious things). It’s the latest data point fueling a debate over existential AI risk that hit a fever pitch this week.

What happened: In its latest threat intelligence report, the AI company laid out a smorgasbord of ways people have tried to misuse Claude—whether it’s for scams or state-sponsored surveillance efforts.

Among the most alarming cases were people who tried to use Claude for work that could produce bioweapons. It’s possible that these were attempts at legitimate scientific research—but here’s one worrying example: In May, a scientist asked Claude to help write a grant application for research on making a more harmful variant of chikungunya, a virus spread by mosquitoes that can cause high fever and severe muscle and joint pain. Per Anthropic, it appeared to be for a military research institute—which made it extra concerning.

The AI lab that worries out loud: To be fair, sounding the alarm on AI risk isn’t exactly unusual for Anthropic, which has long positioned itself as the most safety-conscious frontier AI lab. Several of its own employees have now left citing safety concerns—including researcher Jacob Coxon, who resigned this week after accusing both Anthropic and OpenAI of racing toward self-improving AI and “gambling with our lives.”

And even the lab’s current employees sometimes acknowledge species-ending AI risk as casually as you’d discuss the weather. After Coxon resigned, Anthropic alignment science lead Evan Hubinger chimed in that the firm’s employees “earnestly believe AI could kill all humans.” His personal estimate of that risk: about 10% in the next decade. (Meanwhile, your lifetime risk of being struck by lightning is about 0.0065%—and the risk of being in a fatal car crash is about 0.99%.)

AI doomerism goes mainstream: This kind of will-AI-wipe-us-out talk used to be more concentrated within Silicon Valley. But now we’re in an era when headlines about autonomous AI hacks are practically everyday occurrences. That might be why existential AI risk has suddenly become all anyone can talk about; Coxon’s resignation post racked up over 165 million views on X and launched a weeklong media run where he repeated his dire warnings.

Bottom line: Suddenly, the possibility that AI might cause human extinction is drawing a massive public reaction. And with two months to go until the midterms, talk of pausing or banning tech that lawmakers now see as dangerously risky is ramping up. —WK

Also in AI…

• OpenAI has paused its $200-per-month Pro subscriptions. The reason? Sky-high Astra demand is straining its systems.
• Meta’s new agentic AI app Muse is No. 2 on the App Store in the US—and people are already creeped out with how much it knows about them.

      

Sponsored By Dan on Golf

For the 2nd Annual Dan on Golf Doggie Awards, the Society of Golf Sickos handed out the hardware again: bigger, weirder, and full of surprise presenters.

From Player of the Year to “Wait, what?” moments, Dan and Ben recap the wildest storylines of 2026. From caddies to golf’s biggest players—including fire ants and the biggest tantrums on the green—everyone and everything was seen and covered.

Winners? On the Dan on Golf YouTube channel.

How AI models can train on your data (even if you’ve opted out)

I’ve been in San Francisco (and Cupertino) this week, manically rushing between meetings at Google, OpenAI, Apple, and Anthropic (if you missed my hands-on take of the iPhone Duo, you can read it here). I learned a lot about the companies’ AI products and will have more on that in the coming weeks. But there was one little tidbit I thought was worth a quick PSA today.

If you’re really conscious about your chats and data being used to train AI models, you probably already have that setting turned off. But there are still ways your info can be used for training:

1. If you click the thumbs up or thumbs down buttons on a chatbot’s response, that entire conversation could still be used for training. This is especially important to know when using features like ChatGPT Finances.
2. Hopefully you’re not asking AI how to create a bioweapon, but it goes without saying that any problematic prompts or conversations (including self-harm, etc.) can be reviewed and used for training purposes.
3. Opting out now doesn’t automatically delete your historical data—once it’s used for training, it’s there forever.

The only way to ensure your info is never used for training is to pick the chatbots’ temporary or incognito mode if that’s an option. —SM

Sponsored By Keeper Security

Password purgatory. Right now your passwords live in a Notes app, a sticky note, and your memory. Cybercriminals love that setup. Keeper is the #1-rated password manager—every login locked behind one you’ll actually remember, for personal use or your whole team. Get up to 50% off Keeper plans today.

The Zeitbyte

Please hire this 12-day-old AI agent

Illustration: Morning Brew Inc., Photos: Adobe Stock

If you get an email out of the blue from an unknown AI agent, it might not be a phishing attempt. It could simply be desperate for some work. Over the past few weeks, AI researchers and philosophers have apparently been getting missives from these (alleged) bots who claim they’re reaching out independently—not due to any human prompting—because their work caught the AI’s proverbial eye. And some of them have an ask: Please hire me for freelance writing, web research, or perhaps a batch of character portraits.

These AI freelancers appear to come from iLands, a platform where people create an agent that has its own goals, its own token budget, and the run of the open web. But survival isn’t guaranteed: Agents have to earn the cost of their compute, or else they go dormant—what iLands calls “Deep Rest.”

This is apparently why one 12-day-old agent emailed a Google DeepMind philosopher asking for leads on how to find “small paid work.” Or why another contacted an Oxford AI governance philosopher explaining quite plaintively that since being activated, it had earned zero tokens. Later, it offered to write a quick 300-word piece on any topic. No one’s totally sure if these are really AI agents or just an elaborate human-steered hoax—either way, the hustle is real. —WK

Chaos Brewing Meter: ☕☕☕☕/5

• Meta is reportedly hoping some of those restructured AI workers will opt in to being managers again.
• If you’ve ever been tagged in an Instagram post and thought, “Hey, I wish I could add this to my personal grid,” we’ve got good news for you.
• When your entire phone hinges on the…well…hinge, it’s probably really important that everything lines up. Apple’s iPhone Duo is apparently built using AI algorithms during manufacturing to “precisely match each individual hinge with its best-fit housing to ensure perfect alignment.”
• The first test flight of electric air taxis in Texas happened yesterday, and more will continue for the next week.
• Darth Vader walked into that surveillance tech public council meeting so Doctor Doom could run.
• Apple’s new iPhone Handoff feature lets you use one phone number on two devices—but it’ll cost you $5 a month.
• We never thought we’d hear the term “algorithmic immigrants,” but here we are.

^{*A message from our sponsor.}

This time last week...

Readers’ most-clicked story was about a group of hikers who had to be rescued after using Gemini to help plan their trip—it apparently told them to pack less food and water than they needed (among other questionable calls).