Google's Robot Brain Shipped With a No-Go List

So, What Actually Happened?

So I opened a robotics launch expecting benchmarks and found a list of places the model is not allowed to go. Google shipped Gemini Robotics ER 2 and told developers not to use it in healthcare or transportation, or anywhere a malfunction could hurt somebody. Then Germany quietly handed AI Act monitoring to its telecom regulator, which is the dullest sentence written this week and probably the most consequential. We scanned 190,000 articles this week so you don't have to. Meanwhile a self-spreading exploit riding a writing assistant survived two patches and is still live. Three different desks, one thing happening underneath.

The Bottom Line: Your AI vendors have started writing down what their systems must not be used for. Read those lines before your lawyers have to.

The Tracks That Matter

1. An AI Worm Survived Two Patches and Is Still Live

A self-propagating exploit riding Copilot for Word is still live after two patches, which should be ruining somebody's quarter right now. The mechanics matter less than the ownership question. A worm that travels through the documents an assistant reads and writes does not sit inside one product boundary, so each patch closes a symptom while the propagation path stays open. The security-policy trade spent the same week arguing that acceptable-use rules need technical enforcement rather than a signed PDF, which is the same finding stated politely. Your AI acceptable-use policy is not a control. It is a document describing a control you may not have built.

Here's what works: Ask your security team one question this week. If an AI assistant became the delivery path, which specific control stops it? Name the control, not the policy.

2. Germany Hands AI Act Enforcement to Its Telecom Regulator

The Bundesnetzagentur, the agency that already regulates German telecoms and power grids, takes the central role in monitoring AI regulation. No new agency, no new building, just an existing regulator with existing enforcement muscle pointed at a new law. That is what implementation looks like when a rule stops being a rule and becomes somebody's job description. Across the Atlantic the argument is a step behind: Brookings spent the same week making the case that Congress must pass a federal law on AI governance at all. One side is naming the desk that will call you. The other is still drafting the reason to call.

Here's what works: If you sell or operate in the EU, find out this week which national authority covers your sector. That is the letterhead your first compliance letter arrives on.

3. Google Ships a Robot Brain With a No-Go List

Gemini Robotics ER 2 landed with real capability. It reasons over live video, plans multi-minute tasks, hands work between different robots, and calls tools mid-task to move actual hardware. Then the model card draws a line: not for safety-critical work, naming healthcare and transportation and anywhere a malfunction could foreseeably cause death or injury. A companion safety benchmark tests whether the model refuses unsafe tool calls and asks a human when it is unsure. Meanwhile the data platforms are pitching agents on production lines making trusted decisions in real time. Somebody has to reconcile those two documents, and it will not be the vendor.

Here's what works: Pull the model card for every AI system touching physical operations. The prohibited-use section is a contract term your vendor already wrote for you.

Quick hits:

• Bloomberg is buying its way into private-markets data. The Canoe Intelligence purchase drops alternative-asset document automation inside the terminal, which tells you where the auditability money is going.
• A billing vendor exposed 1.26 million patients and took its time saying so. The delayed disclosure is the real story: your third party's clock is your clock.
• China's open-weight lead is now a US policy problem. CNBC laid out the open-model policy gap, with Together AI's chief executive arguing that a company's own data is the strategic asset it should stop shipping to closed providers.

Signal vs. Noise

🟢 Signal: who signs off on AI. Governance, risk management and compliance all gained real ground across the industry this week, with compliance climbing hardest among the words that actually attach to a decision. Most coverage is still scoring model capability against model capability, which is not the question any buyer with a legal department was asking.

🔴 Noise: ”automation” as a word. It held heavy volume across the wires while losing its grip on what is actually being decided. Nobody is arguing about whether to automate anymore. They are arguing about who answers for it, and ”automation” does not carry that meaning.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Google published a list of places its new robotics model may not go, Germany named the regulator who will check AI Act compliance, and a self-spreading exploit in a Microsoft writing assistant survived its second patch, all inside the same 48 hours.

Read alone, each lands on a different desk. The robotics press covers a capability launch. The Brussels trade covers an administrative appointment. The security wires cover a patch that did not take. Put them on one morning and they describe the same movement: liability is being written down. A vendor naming forbidden uses, a state naming an enforcer, and a live exploit nobody has claimed are three points on one line, and that line runs through your procurement file. For two years the working assumption in enterprise AI was that responsibility would get sorted later, probably by the vendor, probably in a contract nobody read. Later arrived this week, and it arrived as text.

What changes on Monday is boring and it is the whole job. Open the documentation for every AI system you run and find the section describing what it must not be used for. If that section is empty, you are not holding a safer product. You are holding the liability your vendor declined to write down.

By The Numbers

• 1.26 million patients exposed in a billing-vendor breach — the notification came late, and the delay is the part your auditors will ask about.
• groundcover raised $100 million in a Series C — capital keeps flowing to systems whose job is watching other systems run.
• Bolt.new is running $1.3 million in revenue per full-time employee — the efficiency benchmark MIT Sloan says investors now measure your last round against.
• Apollo 2 picked objects off the floor 45.7% of the time — the honest gap between a robotics demo and a shift you could actually staff.
• P-1 AI closed a $50 million Series A led by NEA — engineering-design AI is pulling frontier-lab names onto the cap table.
• See what's rising in our 190K-article corpus this week →

Deep Dive: The Rider

Every touring DJ has a rider. Two pages saying what has to be in the booth, what the venue owes you, and buried near the back, what you will not do. No opening for that act. No playing past two if the limiter is on. It is not diva behaviour. It is the only document in the building that says whose fault it is when the night goes wrong.

The rider moved into the model card
A capable robotics model shipped this week with a list of where it must not go: healthcare, transportation, anywhere a failure could injure someone. Vendors spent two years selling capability with the limits left unspoken. This week the limits got typed, and typed limits behave differently than assumed ones.

The venue named a safety officer
Germany handed AI Act monitoring to an existing regulator with existing teeth. Brussels rules stop being abstract the moment a specific agency owns them and has a phone number. The US is a step behind, still arguing whether the venue needs a licence at all.

The gear still catches fire
Meanwhile a worm in a writing assistant survived two patches, and a billing vendor sat on a breach touching 1.26 million people. Riders and safety officers do not stop equipment failing. They decide who pays for the repair, and they decide it in advance.

What Actually Works

1. Read the prohibited-use section first: it is the part of a model card written by lawyers, which makes it the part that binds.
2. Name your national authority: for every EU market you touch, one agency now owns your sector. Find it before it finds you.
3. Treat assistants as network reach: your AI tools read and write documents across every team. That is a delivery path, not a feature.
4. Put a number on third-party disclosure: contract the hours, not the intention. A delayed breach notice is what ”promptly” means when nobody wrote a figure.

Nobody reads the rider until something goes wrong. That is precisely when it becomes the only page in the room that matters.

What's Coming

Every Member State Names Its Desk

Germany handing AI Act monitoring to its network agency is the template, not the exception. Expect the rest of the bloc to appoint existing sector regulators rather than build new ones, which means your AI compliance contact will be the regulator you already know from telecoms, finance or medical devices.

Model Cards Become Procurement Documents

Once a vendor writes down where its model must not be deployed, that paragraph starts showing up in RFPs and insurance questionnaires. The awkward consequence: vendors who publish no boundary at all will start to look less safe, not more.

Disclosure Timing Becomes the Audit Question

The 1.26 million patient records exposed through a billing vendor landed with a late notification, and the lateness is the reviewable failure. Expect disclosure-clock language to arrive in vendor contracts well before the breach math changes.

For Your Team

Monday's meeting prompt: ”If one of our AI vendors published a list tomorrow of uses their model is not approved for, and one of our live workflows was on it, who in this room finds out, and how long does that take?”

Share-worthy stat: A robotics model shipped this week that reasons over live video and coordinates multiple robots. The same release told developers not to use it in healthcare or transportation. Capability and disclaimer arrived in one document, and most buyers are only reading half of it.

Go deeper: Track where AI governance and accountability decisions are moving, in real time →

The Track of the Day

”Google tells developers not to use the robotics models for safety-critical work, naming healthcare and transportation, or anywhere a malfunction could foreseeably cause death, injury or property damage.”
— Gemini Robotics ER 2 model card

A capability launch and a liability notice in the same file. Read both, because your vendor has already decided which one they will point at later.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: July 31, 2026 | Curated by Yves Mulkers @ Ins7ghts