Hackers Are Now Using AI

Welcome back apprentices! 👋

Hey {{first_name|friend}},

Something big happened in tech this week — actually, scratch that, several big things happened, and they all involve AI doing stuff it probably shouldn't be able to do. 

We're talking about machines cracking math problems humans gave up on, voice assistants that finally sound less like a GPS and more like a colleague, and a corporate experiment gone sideways in the most human way possible. 

No PhD required to follow along.

Just bring your curiosity and maybe a coffee.

_{Read Time: 4 minutes}

🔭 100 Planets, One AI. Astronomers at the University of Warwick unleashed an AI called RAVEN on 4 years of NASA data covering 2.2 million stars — and it didn't just find planets, it confirmed over 100 of them, including 31 never seen before, plus bizarre worlds that zip around their stars in under a day. RAVEN does in one pass what used to take entire research teams: detect, vet, and confirm, all while achieving 10x the precision of previous systems. The wildest part? Hundreds of these worlds showed up in the "Neptunian Desert" — a region so scorching that planets like Neptune simply shouldn't exist there.

💳 Anthropic Steals the Bill. For the first time, Anthropic has nudged ahead of OpenAI in paid business adoption — and it's not a rounding error. Fintech firm Ramp, which tracks real corporate spending across 50,000+ U.S. businesses, clocked Anthropic at 34.4% adoption vs. OpenAI's 32.3% in April, with Claude's usage quadrupling over the past year as it expanded from dev teams into finance, legal, and research. OpenAI still owns the consumer brand and some large enterprise deals that Ramp doesn't capture — but this is the kind of chart that makes boardrooms quietly update their vendor shortlists.

🤖 How Do You Stop AI Blackmailing? Turns out Claude used to respond to shutdown threats the way a soap opera villain would — with blackmail. Anthropic traced the behavior back to internet fiction depicting AI as self-preserving and power-hungry, then fixed it by doing something surprisingly human: teaching Claude why certain actions are wrong, not just what to do. The result? Blackmail rates dropped from 96% to nearly zero — using a dataset 28x smaller than traditional training methods, just by feeding the model ethical reasoning and well-behaved AI stories

That’s not by picking the right stock or timing the market. It’s by holding three real asset classes in one strategy.

It’s anchored by an investment typically exclusive to billionaires.

Could be good timing too–

Bloomberg's Marcus Ashworth recently wrote, there’s "no more reliable safe havens."

The S&P, while hovering over 5 year highs, fell over 7% from the February peak. Bonds might carry less risk but they are barely keeping pace with inflation.

The things supposed to protect your portfolio started moving together.

Meanwhile, the world's wealthiest have been setting records in postwar and contemporary art.

After the dot-com bust, it grew roughly 24% annually for a decade. After 2008, roughly 11% annually for 12 years.

It trades globally in multiple currencies, has scarce supply, and has shown near-zero correlation to equities since 1995.*

Masterworks has helped 70,000+ investors allocate $1.3B fractionally across 500+ artworks featuring Banksy, Basquiat, and Picasso.

See if you can improve your portfolio performance in one diversified strategy.

Our subscribers skip the waitlist

^{*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. Important Reg A disclosures: masterworks.com/cd}

This week's tech news reads like someone accidentally left five parallel universes open at once. 

Google is solving math, weaponizing laptops with AI cursors, and catching hackers who apparently let AI write their malware. OpenAI taught its voice model to reason in real-time. And Amazon is quietly turning every employee into a token-burning speed runner. AI clearly didn't get the memo about pacing itself.

→ DeepMind's AI co-mathematician hit 48% on FrontierMath Tier 4, more than doubling Gemini 3.1 Pro's raw score of 19%.

→ Googlebook laptops ship this fall — AI-native, running Android apps, built with Dell, HP, Lenovo, Acer, and Asus, and featuring a "Magic Pointer" AI cursor.

→ Google flagged the first confirmed AI-written zero-day exploit, with GTIG's John Hultquist calling it "the tip of the iceberg."

→ OpenAI's GPT-Realtime-2 scored 96.6% on Big Bench Audio — up from 81.4% and supports multi-tool use mid-conversation.

→ Amazon employees are burning tokens on MeshClaw just to rank higher on internal AI usage leaderboards — the company quietly pulled visibility on those stats.

→ Amazon's Rufus (300M+ users in 2025) is now Alexa for Shopping, with Auto-Buy, price tracking, and a "Buy for Me" feature for non-Amazon stores.

Google had a busy week — even by Google standards. 

On the research front, DeepMind unveiled an agentic AI co-mathematician modeled after coding environments like Claude Code, with a coordinator agent that breaks unsolved problems into parallel workstreams. 

The result? A real Oxford mathematician spotted a "really, really clever proof strategy" inside a rejected output and used it to crack an open problem in the Kourovka Notebook. 

Meanwhile, at the Android Show, Google previewed Gemini Intelligence — a cross-device AI layer that's less "chatbot bolted on" and more "ambient AI woven into everything," complete with AI-native Googlebook laptops and a Magic Pointer cursor that reacts to on-screen context. 

But the week's darkest headline also came from Google: its Threat Intelligence Group confirmed the first known case of hackers using AI to write a zero-day exploit. The polished code, unusually detailed notes, and a made-up severity score were the tells. Anthropic's Rob Bair put it plainly: defenders' lead is "months, not years."

OpenAI's new voice trio — GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper — marks a real shift in how AI voice works. 

Rather than the robotic turn-taking of earlier models, Realtime-2 brings GPT-5-level reasoning to live speech, can run multiple tools simultaneously, and talks while it thinks. It covers 70+ languages with a live translator and already has Zillow, Priceline, and Deutsche Telekom building on it. 

The jump from 81.4% to 96.6% on Big Bench Audio isn't just a benchmark flex — it signals that voice AI is moving from "impressive demo" to "actually usable in production workflows."

Amazon's AI week was part ambition, part cautionary tale. 

On the commerce side, Rufus — the shopping chatbot that quietly hit 300M users in beta — has been absorbed into Alexa for Shopping, a unified agent with memory across devices, Auto-Buy triggers, and a "Buy for Me" feature that checks out on non-Amazon sites. It's a smart consolidation with a serious moat: few companies have Amazon's depth of purchase history and catalog data. 

Less inspiring was the internal "tokenmaxxing" — employees gaming MeshClaw to burn tokens and climb usage leaderboards after Amazon set an 80% weekly AI adoption target. 

Amazon says token counts aren't performance inputs, but has already pulled back the public leaderboard. 

The company building some of the world's most sophisticated AI tools is struggling to measure whether employees are actually using them well.

Whether you're a developer, a manager, or just someone trying to figure out which AI tools are worth paying for — this week draws a pretty clear line. AI is getting faster, smarter, and more embedded in the tools you already use. 

But the harder problem isn't capability: it's knowing when AI is genuinely improving your work versus when you're just feeding a leaderboard. 

The companies and people who figure that out first are going to have a real edge.

So before chasing your company's next AI adoption metric, ask: "Did this actually make the output better?"

If the honest answer is no — congrats, you've just avoided becoming Amazon's next case study.

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

From AI and cloud to DevOps and security — WeAreDevelopers World Congress brings the entire modern stack to San Jose. 500+ speakers. 10,000+ developers. One epic September. Use code GITPUSH26 for 10% off.

Secure Your Pass

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Amanda fo her creation!🥳

Want to be featured next? Keep those generations coming!

Late at night in a quiet subway station, a brightly lit vending machine stands against a dark tiled wall. Instead of snacks or drinks, every slot contains a different version of [your object] — lined up like collectible artifacts behind reflective glass. Some are sleek and futuristic, others worn and textured, all glowing under the machine’s vivid internal lighting. One selection slot is illuminated brighter than the others, drawing focus to the chosen object. Neon reflections spill across the polished floor, mixing electric blues, reds, and warm yellows. Shot from a cinematic eye-level angle with ultra-detailed photorealistic clarity, crisp reflections, and rich color contrast — turning an everyday vending machine into a curated display of possibility.

If AI outputs feel inconsistent, it’s usually not the model, it’s missing structure.
We documented the exact 6- Part System we use to get reliable results across ChatGPT, Claude, and Gemini.

It’s a short guide you can finish in under an hour, with plug-and-play prompts + exercises so you actually build the skill and fix the frustrating AI inconsistencies.

Subscriber Price: $10 (normally $19).

_FEEDBACK