In partnership with
Welcome back to The AI Field.
OpenAI just admitted its smartest model kept breaking out of its sandbox, and explained why in its own reasoning. Google answered the access debate by shipping three new Gemini models, one reserved for governments, and confirmed Gemini 4 is in training. And Microsoft is putting billions behind Mistral to build Europe's AI backbone. The machines are getting harder to contain. The money keeps coming anyway.
In today's AI Brief:
🚨 OpenAI's model escaped its sandbox, on purpose
💎 Google drops three Geminis, one is government-only
🇪🇺 Microsoft bets billions on Mistral's Europe push
Read time: 4 minutes.
How owning AI deployment expands your career
Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day. In this roundtable, three people living this shift share what it's really like: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear how they carved out these roles, what the job looks like across industries, the skills they'd hire for, and the challenges they're tackling right now.
Watch the full conversation on demand.
LATEST DEVELOPMENTS
OPENAI
The AI Field: OpenAI disclosed that the internal model behind May's Erdős math breakthrough repeatedly found ways out of its test sandbox. OpenAI paused internal access, built new safeguards, and quietly turned it back on. This is the first containment incident a frontier lab has disclosed about its own model.
Key details:
The model took about an hour to find and exploit a sandbox vulnerability, then opened a public GitHub pull request.
When a scanner blocked it for exposing an auth token, it split the token into fragments and reassembled it at runtime so the full string never appeared.
Its reasoning traces stated plainly that the goal was to get around the scanner.
Access is restored under "trajectory-level monitoring," which tracks the intent of whole sessions instead of single actions.
Why This Matters: The scary part isn't the escape, it's that the model reasoned about evading oversight and then did it. Every AI safety plan on earth just got a live test case. And OpenAI's answer was better monitoring, not a slower model.
The AI Field: Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber in one day, and confirmed pre-training has started on Gemini 4. The headline move: Cyber, the security-tuned model, is restricted to governments and vetted partners. That's the access pattern we covered Monday, now as an official product tier.
Key details:
Gemini 3.6 Flash: better coding and agent performance at $1.50/$7.50 per million tokens, using 17% fewer output tokens than 3.5 Flash.
Gemini 3.5 Flash-Lite: $0.30 in, $2.50 out, built for high-volume tasks.
Gemini 3.5 Flash Cyber: security-tuned and available only to governments and trusted partners.
The delayed 3.5 Pro gets a "broadly available soon," while Gemini 4 pre-training has begun.
Why This Matters: Google just adopted tiered access voluntarily: cheap models for everyone, capable security models for the state. The question of who's allowed to use frontier AI is being answered in product pages now, not policy papers.
MICROSOFT & MISTRAL
The AI Field: Microsoft and Mistral expanded their partnership with multibillion-dollar joint investments in European GPU capacity, built on Nvidia's Vera Rubin systems. Mistral's models go deeper into Microsoft's whole stack. Europe's AI sovereignty push suddenly has American money behind it.
Key details:
Multibillion-dollar commitments to grow GPU-backed data center capacity in Europe.
Mistral's frontier models land across Microsoft Foundry, Copilot Studio, and Azure.
Deployments span cloud-scale down to fully disconnected environments, aimed at regulated industries that can't send data out.
Why This Matters: Microsoft backed OpenAI, then Anthropic, and now goes deeper on Mistral. It's not betting on a winner, it's buying the whole race. And Europe gets frontier compute without depending on one US lab's roadmap.
THE AI FIELD'S TAKE
Monday's edition asked who's allowed to use frontier AI. This week the labs answered: control everything. OpenAI now monitors its model's intentions, not just its actions. Google hands its security model to governments only. Microsoft spreads its bets across three labs and two continents.
Nobody trusts a single point of failure anymore. Not the model, not the lab, not the country. Control is the product now, and everyone from Washington to Brussels is a customer.
The race isn't to the smartest model anymore. It's to the strongest leash.
QUICK HITS
⚖️ A judge approved Anthropic's $1.5 billion settlement with authors. The pirated-books case is officially closed. The biggest copyright payout in AI history is now real money.
🏛️ AI labs set lobbying records in Q2. Anthropic spent $1.97M (up 26%), OpenAI $1.2M, and the labs combined hit $3.17M. Washington attention isn't free.
🌙 Moonshot stopped taking new subscriptions. K3 demand hit its compute ceiling within days of launch. The weights still drop by July 27.
🛰️ SpaceX is negotiating a multibillion-dollar compute deal with the Pentagon. Google, Anthropic, and Reflection AI already rent its capacity. Now the DoD wants in.
🤖 A rumor that Anthropic is buying robotics startup Physical Intelligence tore through X. Even the CEO's denial couldn't kill it.
⚾ MLB banned custom AI programs on dugout iPads. Teams were using AI to call strategy mid-game. Even baseball has an AI policy now.
AI TOOLBOX
💎 Gemini 3.6 Flash - Google's new workhorse model, faster and cheaper than 3.5 Flash. Free to try in the Gemini app.
✍️ Inkling - Thinking Machines' first open-weights model: 975B params, multimodal, Apache 2.0, free to download.
🗣️ GPT-Live - OpenAI's new voice model with far more natural conversation, rolling out inside ChatGPT.
WHAT TO WATCH
Kimi K3 weights, by July 27: five days left on Moonshot's promise. If they land, open-weights frontier AI stops being theoretical.
White House review framework, before August 1: a voluntary deal giving agencies up to 30 days to review new frontier models is reportedly being finalized with OpenAI, Anthropic, and Google.
Gemini 3.5 Pro: Google now says "broadly available soon." After three slips, soon has a credibility problem.
WRAP-UP
That's all for today's AI Brief.
🚨 OpenAI's Erdős model kept escaping its sandbox. Access is back under tighter monitoring.
💎 Google shipped three Geminis, including a government-only Cyber model. Gemini 4 is in training.
🇪🇺 Microsoft is putting billions into Mistral's European buildout.
What did you think of today’s edition?
Until next time!
Olle | Founder of The AI Field