Welcome, humans.

Big news: we are re-launching our podcast The Neuron: AI Explained with the first new episode set to come out this Friday! If you’re an advertiser looking to sponsor this type of content, reach out to [email protected] and let us know.

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

• Cyber crimes are escalating at a rapid clip thanks to AI.

• OpenAI's o3 model set industry records on ARC-AGI at a fraction of competitors' cost.

• AI coding agent capabilities now basically double every four months.

• AI systems surpassed 94% of human experts in practical virology knowledge.

Cyber criminals are going WILD with AI (because duh?)

By now you’ve probably seen the viral trend of everyone generating action figures and dolls out of their own likeness (it’s SO viral, people are already sick of it). 

Well, as it turns out, giving AI info about your personal interests, hobbies, and even your likeness can be used in social engineering and scams—ESPECIALLY if you share the pics on social media, where bad actors can scoop those identifiers up to use against you.

A recent report out of the UK found that AI-enabled crime, such as phishing and romance scams, have seen a “substantial acceleration” thanks to AI making cybercrime easier and more effective than ever. Wow, who ever could have predicted this?!

But how? AI lowers the bar for criminals, letting them…

• Automate attacks at scale.

• Generate hyper-realistic fake content (websites, reviews, emails).

• Use deepfakes or voice cloning to impersonate trusted people.

For example, remember the exec who was tricked by a deepfake CFO into sending someone £20M? That’s the power AI gives scammers.

Microsoft just reported thwarting $4B in such fraud attempts last year alone.

It's not just phishing and fake websites either. AI fuels convincing job scams with fake AI interviews and automates the manipulation in romance fraud (like pig-butchering scams).

There’s even a new threat for developers: “Package Hallucinations,” where code-generating AI recommends non-existent software packages. Attackers then publish malicious code under those fake names, tricking unwitting vibe-coders into installing malware (one study found 200K unique fake packages generated by popular AI models!)

While companies like Microsoft are deploying defensive AI and building safer products (“Fraud-resistant by Design”), and researchers are testing mitigations (though often with trade-offs, like reduced code quality), vigilance is key.

Our take: Assume AI is amplifying threats. Verify any job offers, ignore urgent payment requests, be skeptical of DMs, and think carefully about the data you feed AI tools or share online. Staying informed and cautious is your best defense in the age of AI-powered scams. Here’s our full explainer with more findings and tips.

We’ve also read it’s probably a good idea to have a safe phrase with your significant other and/or family members to confirm you’re actually talking to them on the phone when sharing personal details. Why? Check out today’s Treats to Try…cloning voices just got INSANELY easy (as if it wasn’t easy enough as it is). 

FROM OUR PARTNERS

Do real work with agentic AI

Sana’s all-in-one AI platform lets you build and deploy powerful agents that automate complex workflows for every team.

With Sana Agents, you can:

• Conduct 9 hours of deep research in seconds across every app.

• Analyze complex datasets 10x faster.

• Accelerate your sales cycle by cutting prep time 90%.

• Capture perfect meeting notes with instant action items

LLM-agnostic. Extensible APIs. Enterprise-grade security.

Trusted by top teams, loved by end users.

Check it out here

Prompt Tip of the Day

Apparently OpenAI is now secretly watermarking text from their newest models with invisible characters that look identical to regular spaces, but can’t be detected by typical tools. This sneaky feature seems to target longer outputs like essays that students might copy-paste.

Want to check if your AI text is watermarked? Paste it into this online character viewer or open it in a code editor like Sublime Text or VS Code—the watermarks create a distinctive pattern that clearly identifies direct copies from newer ChatGPT models.

Removing them is super easy, too—just find and replace these special characters with regular spaces in any text editor that displays special characters.

Btw, Hacker News isn’t convinced this was intentional on OpenAI’s part, but it’s an interesting discovery either way.

Treats To Try.

1. *Data brokers selling your personal info? Yikes. Incogni Unlimited not only scrubs your data from our extensive 250+ broker network, but lets you point them at ANY sketchy site sharing your details. Protect your personal data for 55% off with code NEURON

2. Instagram Edits is the company’s new competitor to CapCut, and comes with AI tools that Instagram doesn’t have.

3. Sand AI released MAGI-1, an open-source video model that generates high-quality videos chunk-by-chunk with precise timeline control and infinite extension capabilities (demo on X)—try it here.

4. Nari Labs released Dia 1.6B, a new open AI model that converts your text into natural-sounding conversations with emotional tones, speaker variations, and non-verbal sounds like laughter and coughs—demo, try it here.

5. Assort Health answers patient calls with specialty-specific voice AI, freeing your medical staff from routine scheduling tasks (raised $22M)

6. Descript now has an agent to help you “vibecode” videos (so like, “vibe editing”?); it’ll remove awkward silences, translate content, and condense long recordings into concise final products—watch this demo or apply here.

7. Superpower analyzes your blood twice yearly to catch health problems before symptoms appear and creates personalized action plans to improve your wellbeing—paid only rn ($499/year).

8. Manychat automates messaging and responses across Instagram, WhatsApp, TikTok, and Messenger—free to try (raised $140M).

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

Around the Horn.

• OpenAI would apparently buy Google Chrome if given the chance, and may be partnering with Shopify to let you buy Shopify products directly in chat.

• Grok now has vision, allowing it to see the world around you like ChatGPT can.

• Physical Intelligence's π0.5 robot can now clean up in brand new homes it's never seen before by understanding both what to do and how to do it.

• This video is a WILD demo of what AI gaming will be like when NPCs in games can respond to what you actually say—its from the game Whispers from the Star (yet to be released).

• OpenAI’s o3 model now scores 57% on the ARC-AGI v1 test (a tough benchmark for AI intelligence):

  • Mike Knoop shared on X that o3 (medium) is now the industry's leading AI reasoning system with double the score at 1/20th the cost compared to competitors, scoring $1.5/task (o3 high kept timing out).

  • As Ethan Mollick says, that’s more evidence that o3 represents “a big move forward” for AI.

  • That said, it only achieved 3% on ARC v2, at $2.5/task—talk about a challenge!

  • Related: Because of o3/o4, the time horizon of coding agents (tasks measured by how long they would take humans to do) is now doubling every 4 months (up from 7 months reported back in March)—read our full story on the original report here.

Midweek Wisdom

• Anthropic discovered that Claude expresses five main value categories (Practical, Epistemic, Social, Protective, and Personal) in real-world conversations, with the AI sometimes mirroring user values but also standing firm on core principles when users request unethical content (paper)

• Check out this startling analysis showing frontier AI systems have already surpassed 94% of human experts in practical virology knowledge (paper).

• This LLM Roadmap blog is a fun way to learn about all the technical stuff related to AI—but it’s still VERY technical.

• Check out the top AI papers from last week (according to NLP newsletter).

• This is a great piece on how there are two OpenAIs… the one promising superintelligence, and the one that’s just a business.

• Pretty cool: AI is being used to protect endangered wildlife through thermal cameras that alert ships to nearby whales, text message systems warning villagers of snow leopards near livestock, drones counting waterbirds in seconds instead of thousands of hours, and monitoring tools that have detected thousands of illegal wildlife sales across social media platforms.

A Cat's Commentary.