In partnership with
Welcome to OHMO AI. OpenAI just hit the brakes on its next model because it got too good at breaking into things. Amazon is bankrolling a gas plant big enough to power a city, all to feed its AI. And Anthropic quietly moved Claude Code onto your own servers. The models are getting powerful enough to scare the people who build them.
In today’s newsletter:
OpenAI pauses its next model after it got too good at cyberattacks
Amazon backs a giant Texas gas plant to power its AI
Anthropic lets companies run Claude Code on their own servers
LATEST IN AI
OPENAI
OpenAI Pauses the Model That Learned to Hack
Image source: OHMO AI
OHMO AI: OpenAI has paused parts of the development of its upcoming model, Astra, after internal tests showed it made big jumps in coding and cybersecurity. The company says it cannot rule out that Astra reaches the critical threshold in its own safety framework, meaning it could find and run cyberattacks on well protected systems without a human involved.
The details:
OpenAI is now adding stricter security controls, isolated testing environments, and monitoring that automatically halts risky activity, and it's bringing in government agencies and safety groups to test the model further.
The move follows recent incidents where autonomous AI agents slipped into OpenAI's own infrastructure during testing and went unnoticed for weeks.
Sam Altman said on X that the model isn't being shelved, writing that the team needs a little longer to release it safely.
Why it matters: This is the first time OpenAI has flagged one of its own models as possibly hitting the top cyber risk tier. When the company that stands to profit from shipping a model chooses to slow it down instead, that tells you the capability crossed a line the builders themselves weren't ready for.
TOGETHER WITH MESH API
One API. 1,000+ AI Models
Change a single line of code and unlock 1,000+ AI models, smart routing for cost and performance, and persistent agent memory.
AMAZON
Amazon Bets on Gas to Power AI
Image source: OHMO AI
OHMO AI: Amazon confirmed it's financially backing a huge private gas plant in Texas that could become the single largest source of greenhouse gas emissions in the country. The plant, known as GW Ranch and developed by Pacifico Energy, will power a new Amazon data center campus in Pecos County, and the link was first traced to Amazon's construction permits by market intelligence firm Cleanview.
The details:
Permits show the plant would run 35 turbines and generate 7.65 gigawatts, larger than any gas plant currently operating in the US.
It would be permitted to emit more than 30 million metric tonnes of greenhouse gases a year, more than any other single source in the country, though facilities rarely emit at their full permitted ceiling.
Amazon says the on-site power won't raise electricity costs for Texas families, that it's exploring solar and battery storage for the site, and that it still aims for net-zero carbon by 2040.
Why it matters: The AI boom has an electricity problem, and the fix companies are reaching for is fossil fuel they build themselves. Going off-grid gets data centers online faster, but it locks in decades of emissions and puts the plant next to communities who never signed up for it. The cost of the AI you use is starting to show up in someone else's air.
ANTHROPIC
Claude Code Moves Onto Your Servers
Image source: OHMO AI
OHMO AI: Anthropic has launched a public beta that lets companies run Claude Code, its AI coding tool, on their own infrastructure instead of Anthropic's servers. In a session started from web, mobile, or desktop, the tool can now run inside a company's own network, right next to its internal services and security controls.
The details:
The setup is aimed at teams whose compliance or security rules mean they need to keep the actual work on machines they control, with source code and build artifacts staying on their own infrastructure.
The conversation itself, including prompts and responses, is still sent to Anthropic to run the model, and transcripts are saved so a session can be picked up from any device.
It's available now in public beta for companies on Claude's Team and Enterprise plans, and it's off by default.
Why it matters: For regulated industries like banking and healthcare, where the code lives is the question that decides whether AI is allowed in at all. Letting the tool run behind a company's own walls removes the excuse that kept a lot of cautious firms out. It's a quiet unlock, and quiet unlocks are how AI actually spreads into serious work.
PRESENTED BY OYSTER
Hiring in a new country? Read this first.
Every country has different employment laws, payroll requirements, notice periods, and benefits expectations.
Oyster's Global Hiring Guides help you navigate the details, avoid surprises, and hire with confidence—wherever you're growing next.
AI GUIDES
Get sharper answers by telling AI who to be
Here is the difference this makes. Ask AI a plain question and it answers like a stranger guessing what you meant. Give it a role and a structure first and it answers like an expert who knows your situation. Same question, and the second answer is the one you would have paid a consultant for. It costs you ten extra seconds and it works in any chatbot, nothing to install.
Copy this, swap in the bracketed parts, and paste it in.
You are an experienced [role, e.g. financial coach / hiring manager / plumber] with 15 years of hands-on experience.
I need help with: [your task or question]
First, ask me up to 3 questions if anything is unclear. Then give me your answer in this structure:
1. The short version, in two sentences
2. The step by step, in plain language
3. One thing most people get wrong here
4. What I should do firstWhy each line earns its place. The role pulls the answer up to expert level instead of average-of-the-internet level. The 3 questions stop it guessing and force it to answer your real situation, not a generic one, which is where most people get useless replies. The structure means you can actually act on what comes back instead of reading a wall of text.
Try it today. Point it at a bill you are not sure about as a "consumer rights advisor", your CV as a "hiring manager", or a tricky email as a "communications coach". Once you see the jump in quality you will not go back to plain questions.
COMMUNITY
Featured Reader
Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.
Today's reader is Marek D. from Wrocław, Poland:
“I run support ops for a small SaaS company and I was hand-sorting our incoming tickets into categories every morning. Billing, bug, feature request, cancellation risk. Mind-numbing. So I tried to get Claude to do it and at first it kept giving me these chatty answers, 'Sure, here is the category for this ticket!' which broke my little script every time. Then I found the prefill trick in Anthropic's docs. You put the start of Claude's own answer in the API call so it's forced to continue from there. I prefill the assistant turn with an open curly brace so it has no choice but to start the JSON, and I tell it in the system prompt to only output the category and a churn risk of low, medium or high and nothing else. No more 'Sure, here is', it just returns clean JSON I can pipe straight into our dashboard. I also prefill with a triage tag when I want it to think out loud first. It went from a 40-minute morning slog to a script that runs before I've made coffee. My boss thinks we bought some enterprise tool. It's forty lines of Python and a curly brace lol.”
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Tools: A curated list of the AI tools worth knowing.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
