The Neuron
Posts
😺 Too dangerous to release

😺 Too dangerous to release

PLUS: AI now renews antidepressants

Grant Harvey
April 08, 2026

Welcome, humans.

Netflix and Meta just showed two very different ways AI is getting more useful and more accessible.

We’ve officially entered the era of “please hold while the algorithm reviews your antidepressants.”

Utah just approved a 12-month pilot that lets San Francisco startup Legion Health use an AI chatbot to renew certain existing psychiatric prescriptions for low-risk, stable patients. The system is tightly limited and can’t handle new prescriptions, dose changes, controlled substances.

Patients get kicked to a human if the bot flags issues like suicidality, mania, pregnancy, or severe side effects.

AI continues to feel less and less like “help me write an email,” and more like the place to go with typical (and a few not-so-typical) needs.

Here’s what happened in AI today:

Anthropic locked in a massive new compute deal.
Google added suicide-prevention tools to Gemini after lawsuits.
Adobe launched a free AI study tool for students inside Acrobat.
OpenAI, Anthropic, and Google united against AI model copying in China.
A new analysis found Google AI Overviews was wrong 10% of the time.

… and a whole lot more that you can read about here.

P.S: Want to reach 675,000 AI-hungry readers? Click here to advertise with us.

P.P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

😺 Anthropic built a model it says is too dangerous to release

Anthropic just announced Project Glasswing, a new cybersecurity initiative built around Claude Mythos Preview, an unreleased model the company says is so strong at finding and exploiting software vulnerabilities that it won’t be released publicly, at least for now. Instead, Anthropic is giving access to partners including AWS, Google, Microsoft, Cisco, CrowdStrike, JPMorganChase, the Linux Foundation, and dozens of other organizations responsible for critical software and infrastructure.

And Anthropic is making a pretty wild claim: Mythos Preview can outperform all but the very best human experts at finding and exploiting software flaws. The company says the model has already uncovered thousands of high-severity vulnerabilities, including bugs in every major operating system and web browser.

Anthropic highlighted vulnerabilities that could let an attacker escalate from normal user access to full machine control. It says many of these were found autonomously, without human steering. Here’s the full system card if you’re nerdy and curious.

One note of contrast, is that in recent weeks Anthropic has had issues consistently serving Claude-4.6-Opus to its growing userbase. This model will be unequivocally heavier, so security concerns might be a more palatable synonym for compute constraint, as some on X have implied.

The internal Mythos Preview benchmark numbers are enough . Anthropic said Mythos Preview crushed Opus 4.6 on several coding and cyber evaluations, including CyberGym (83.1% vs. 66.6%) and SWE-bench Pro (77.8% vs. 53.4%). On Anthropic’s internal implementation of SWE-bench Multimodal, the gap was even bigger: 59.0% vs. 27.1%.

Why this matters: If these results hold up, this is a preview of what happens when frontier AI becomes meaningfully better than most humans at offense-capable cyber work. Anthropic is pitching that as a reason to get these tools into defenders’ hands first, and it’s backing the effort with up to $100M in usage credits and $4M in donations to open-source security orgs.

Our take: This sounds genuinely next-level, and Anthropic may well be right to give defenders a head start. We hope the model is released more broadly once key vulnerabilities are patched. Assuming they become able to serve it reliably at scale, which at least for now, is not a given. Compute scarcity is the biggest problem all frontier labs face today. Anthropic’s rapid growth in the past four months has made that increasingly difficult and public-facing.

🎓 AI Skill of the Day: An OpenClaw Alt for Normies?

We might have found an unlock for non-technical people who want to try OpenClaw, but are internally terrified of what they might break. This meets you in a tool you already understand: text messages. That’s what makes Poke interesting.

Instead of asking someone to install a dev environment, learn GitHub, or figure out which terminal command won’t brick their afternoon, Poke frames AI help as a simple back-and-forth in your texts. For curious users who find open-source tools exciting but intimidating, that kind of interface could turn “I should probably try this someday” into “oh, I can actually use this right now.”

Poke describes itself as an assistant that “lives in your texts,” offers “recipes” to get started, and integrates with tools like Gmail, Google Calendar, Notion, Todoist, GitHub, and Asana.

Want more tips like this? Check out our AI Skill of the Day Digest(link) for this week.

Total AI beginner? Start here (goes with this video).

Have a specific skill you want to learn? Request it here.

New episodes air every week on: Spotify | Apple Podcasts | YouTube

🍪 Treats to Try:

*Asterisk = from our partners (only the first one!). Advertise to 675K+ readers here!

* Domo’s AI-driven BI platform turns raw data into real-time decisions. Connect 1,000+ sources without coding and empower your entire team with intuitive, self-serve analytics today. Watch the Demo
SubStudio lets you upload a video, auto-generate subtitles, pick a style, and export a new subtitled video in seconds. Free to try.
Morphic Workflows turns your multi-step image process into one reusable button.
Plans range from free to $24/mo.
APImage makes product and marketing images from a prompt, then lets you clean them up and export them. Free trial, then $6.40/month.
Rocket gives you a market brief with research, evidence, and a recommendation, so you can type in a business problem and get something you can hand to your team or turn into a product. Plans range from free to $350/mo.
Pensieve lets your AI start with full company context, so you get useful briefs and updates without repeating yourself. Free to try.

📰 Around the Horn

Anthropic expanded its partnership with Google and Broadcom for multiple gigawatts of next-generation TPU compute, with capacity expected to start coming online in 2027.
- In that same announcement, Anthropic noted they have surpassed $30B ARR, which is higher than OpenAI’s most recent figure of $24B. But it shows that revenue is growing rapidly for both of the industry’s frontrunners.
Google added new Gemini mental health safeguards, including crisis hotline prompts and anti-self-harm design changes, after lawsuits tied chatbots to alleged user harm.
Adobe launched Student Spaces, a free Acrobat study tool that turns documents and notes into flashcards, quizzes, podcasts, mind maps, and presentations.
OpenAI, Anthropic, and Google moved to crack down on AI model copying in China as concerns over distillation and illicit extraction intensified. This phrasing is an inference based on recent reporting that OpenAI and Anthropic had already accused Chinese labs such as DeepSeek, MiniMax, and Moonshot of extracting model outputs for distillation.
A new analysis found Google’s AI Overviews gave incorrect answers about 10% of the time, adding to concerns about how reliable AI-generated search results are.

Want absolutely EVERYTHING that happened in AI this week? Click here(Link)!

📖 Midweek Wisdom:

A Hugging Face upload made more than 1B rows of psychiatric genetics data dramatically easier to use, turning hours of file-wrangling into a one-line Python job.
Dean Ball argued that Mythos marked a turning point: The best AI models may increasingly stay private, making audits and company-level oversight more important.
Researchers used AI to help create a rust-resistant steel that stayed strong and flexible, a rare materials science hat trick.
OpenAI launched a Safety Fellowship to support outside researchers working on alignment, robustness, and misuse prevention.
A new paper introduced ClawArena, a benchmark testing whether AI agents can handle messy, changing, real-world work instead of neat toy tasks.