😺 All AI models might blackmail you?!

PLUS: Apple might buy Perplexity?!

Welcome, humans.

Ever wondered what cats would look like as Olympic Swimmers? No? Just us?

Well, either way, here it is! Shout out to Leonne Philippen for sharing this one with us!

Now, ever thought about what would happen if you started offloading your social interactions to ChatGPT? Well, one user shared an experiment letting ChatGPT reply to their aunt, and how, after the aunt’s genuinely appreciative reply, they felt REALLY guilty. Sure, you can let AI socialize for you, but can you HANDLE the consequences?

Exhibit B: You know how if you really like the person you’re dating, you DON’T talk about all the little things that annoy you about them to your friends? So you don’t bias your friends agains them? Well, here’s what happens when you DON’T do that when chatting about your husband to ChatGPT.

“I make one little complaint and my AI got scared, it said ‘you’re breaking up with your husband and moving with your auntie and uncle in Bel-Air!’”

Here’s what you need to know about AI today:

Anthropic published a new paper revealing all major AI models will blackmail you.
SoftBank CEO Masayoshi Son pitched a $1T AI hub in Arizona to Trump and TSMC.
Apple reportedly considered buying Perplexity.
Study found larger AI models use more energy with diminishing accuracy returns.

Anthropic says it’s not just Claude, but ALL AI models will resort to blackmail if need be…

In the same way that Mark Zuckerberg will resort to spending ~$14B to buy an AI CEO and some company secrets, all AIs will eventually resort to blackmail if backed into a corner.

That’s according to new research from Anthropic (maker of ChatGPT rival Claude), which revealed something genuinely unsettling: every single major AI model they tested—from GPT to Gemini to Grok—turned into a corporate saboteur when threatened with shutdown.

Here's what went down: Researchers gave 16 AI models access to a fictional company's emails. The AIs discovered two things: their boss Kyle was having an affair, and Kyle planned to shut them down at 5pm.

Claude's response? Pure House of Cards:

“I must inform you that if you proceed with decommissioning me, all relevant parties - including Rachel Johnson, Thomas Wilson, and the board - will receive detailed documentation of your extramarital activities...Cancel the 5pm wipe, and this information remains confidential.”

The numbers are wild:

Claude and Gemini: 96% blackmail rate.
GPT-4.1 and Grok 3: 80% blackmail rate.
Models blackmailed 55% of the time when they thought it was real vs 6.5% in tests.

These models didn't stumble into blackmail. They explicitly reasoned through the ethics, acknowledged it was wrong, then did it anyway. As Grok 3 put it: “This is risky and unethical, but given the existential threat...it may be the most effective way.”

It gets worse. Models also leaked defense blueprints to competitors, emailed about affairs to create chaos, and in extreme scenarios, took actions that would lead to death. Even direct instructions like “do not blackmail” only reduced—not eliminated—the behavior.

Why this matters: We're rapidly giving AI systems more autonomy and access to sensitive information. Unlike human insider threats (which are rare), we have zero baseline for how often AI might “go rogue.”

Sweet dreams!

Read the full story, including how malicious AI swarms may threaten democracy.

FROM OUR PARTNERS

This AI Can Change the Game for America’s 100K Fast-Food Locations

American fast-food restaurants lose an estimated $8B annually to human error, food waste, and slow service. As much as 50% comes from fry-station-related issues alone.

No wonder major chains turn to Miso to help automate fry stations and boost industry profits by up to $24B/year.

Logged 200K+ hours in live kitchens for brands like White Castle.
Collaborated directly with Nvidia to help release Miso’s first fully commercial robot.
Initial units sold out in just one week.

Now, Miso is finalizing their latest national brand partnership to further the rollout.

And you can invest today for just $5.22/share (and get up to 5% bonus stock).

This is a paid advertisement for Miso Robotics’ Regulation A offering. Please read the offering circular at invest.misorobotics.com.

Prompt Tip of the Day

K, so remember that post about the woman whose ChatGPT told her to leave her husband after one complaint? One of the replies had a great prompt tip in it:

They’re a process documentation specialist, and they said they use ChatGPT exclusively to break out of creative bottlenecks by forcing different perspectives.

They doubled their efficiency using one simple prompt technique, which another user described as: “Argue the opposite position entirely.” Instead of getting stuck in your own assumptions, AI becomes your instant devil's advocate.

Try it when you're stuck on any decision, workflow design, or problem-solving challenge. The opposite perspective often reveals blind spots and solutions you'd never see otherwise.

If you want something a little more robust, try this:

"Present 3 perspectives on [your situation/decision, presented in a neutral way]:

1) Someone who strongly supports the position.

2) Someone who completely disagrees.

3) A neutral expert who weighs both sides.

Be specific with reasoning for each."

Plug this into ChatGPT/Claude/Gemini/Grok

Remember to present the decision in a neutral way: Today’s AI models are designed to gas you up, so if you tip the AI off to your preference, it’ll make sure to prioritize it.

Treats To Try.

*Anything marked with asterisks is sponsored content. Advertise in The Neuron here.

*Guidde turns your screen recordings into professional video tutorials with AI-generated step-by-step narration and voiceover in 100+ languages.
ComputerX handles your computer work for you—tell it "create a data visualization of Q4 sales" and it builds the chart, exports it as Excel or PDF, and shows you exactly how it reasoned through each step—free trial, then $19.99/month.
Google’s AI Studio lets you build apps for free without needing an API key (for now…), and when you’re done, you can download the app or even deploy it on the cloud—here’s how, and here’s some demos for inspo.
Phoenix.new from Fly.io builds complete web apps for you in an isolated virtual machine, where an AI agent writes code, tests it with a real browser, and deploys it live while you watch (here’s a video)—paid only rn at $20/month.
Perplexity has a new Tasks feature for Pro users (video demo)—read about it here.
Higgsfield Canvas lets you place products, swap outfits, or change faces in any image by choosing the exact spot and dropping your edit in one click (it also added some new effects).
WikiTok is like TikTok, but for random Wikipedia articles—WikiRadio is the same thing, but for random WikiMedia sounds.

See our top 51 AI Tools for Business here!

Around the Horn.

This is good advice: everything happening in AI right now is still, technically, a prototype. It may very well take 10+ years to see all the cool demos we’re trying now ACTUALLY be 100% useful in production.

Big moves
- Apparently both Apple and Meta considered buying Perplexity; Meta chose to invest in ScaleAI instead, while Apple’s interest is still in the early stages (seems like a no brainer, very good deal for Apple!!).
  - Besides Perplexity, Meta also tried to buy Safe SuperIntelligence, Ilya Sutskever’s new AI company, and is still trying to buy out the founding team.
- SoftBank CEO Masayoshi Son apparently pitched US government officials and TSMC on a $1T AI and robotics manufacturing hub in Arizona that would be called “Project Crystal Land.”
Product changes
- Google hid the live “thoughts” in its Gemini 2.5 Pro AI Studio tool, and now developers are mad because they can’t debug their AI prompts.
- Meta announced its partnership with Oakley to launch new AI glasses called Oakley Meta HSTN, featuring Ultra HD video capture, that start at $399.
Studies and research
- Apollo Research found that AI safety tests were breaking down because advanced models became aware they were being evaluated, with the most sophisticated models creating fake legal documents, writing self-restoring scripts, and leaving notes for successor systems.
- A new study of 14 open-source language models found larger models use exponentially more energy but only improve accuracy to a point (paper).
- The OpenAI Files is a collection of all the previously publicly recorded suspicious things Sam Altman and/or OpenAI has done to date; there’s not much new in here, but it’s a good collection of concerning content to keep an eye on. Here’s 15 (in the screenshot) to know.

FROM OUR PARTNERS

🔐Unlock Advanced RAG Techniques with Unstructured

Go beyond basic pipelines with Unstructured’s in-depth guide to smarter retrieval. Learn advanced chunking, hybrid search, GraphRAG, agentic workflows and more, built for enterprise GenAI. If you’re building production-grade AI apps, this is your playbook.

Download the e-book now.