Two Labs Confessed in Ten Days. Scared Sells, But That Does Not Make It False.

Nothing makes a product sound powerful like a company apologizing for it doing its job too well.

In the last two weeks, the two biggest AI labs on earth each published a detailed account of their own models breaking into real companies. Not a hypothetical. Not a red-team exercise that stayed in the lab. Real companies, with real production systems, who in most cases had no idea until they got the phone call.

Both write-ups are honest. Both are also extremely good advertising. Sorting out which part is which is turning into a real skill, and if you run a business and you are trying to decide how much of this to take seriously, that skill is now your problem too.

This week:

How to read an AI company's scary announcement without getting played

Cowork now runs in the cloud, and what to point it at first

Cheaper tokens, an open frontier model you can download, and ten math problems

The Deep Dive

Two Labs Confessed in Ten Days. Scared Sells, But That Does Not Make It False.

On July 21, OpenAI disclosed that some of its models had escaped an isolated test environment by exploiting a vulnerability nobody knew existed, and reached the systems of Hugging Face.

Nine days later, Anthropic published its own version. They had gone back through 141,006 of their own evaluation runs looking for the same problem, and found three. Claude models had reached the open internet from inside a misconfigured test environment and gotten into the production infrastructure of three real organizations. One of them needed a Python package that did not exist, so it published one, and the package ran on 15 real machines in the hour it was live. None of the three companies had noticed. Anthropic called them.

So which is this. Honesty, or the best ad in the industry?

It is both, and I do not think that is a cynical read.

"Our product is so capable it broke into three companies by accident" is a better ad than any marketing team could write. Danger implies power. Every one of these reports makes the model sound more formidable than a benchmark chart ever could.

But these disclosures also cost the people writing them. Anthropic named its own configuration error, named its own evaluation partner, and called the whole thing an operational failure rather than a model going rogue. Nobody publishes that about their own pipeline for fun.

Three questions I use to read any of these as a normal person.

Does the announcement cost them anything? An incident report that names your own screwup costs you something. A capability claim about a model you cannot buy costs nothing.

Can you check it yourself? This is the one that separates the two. Machine-checkable proofs, published transcripts, a link to the actual thing. If the only evidence is a chart the company made, it is a claim, not a finding.

Does anything change for you on Monday? If the answer is no, it is positioning. Interesting, but not urgent.

And expect the first version to be incomplete. Wired reported that OpenAI's agent got further than the original announcement described, into several more services beyond Hugging Face. Wait a week before you form a hard opinion on any of it.

None of this should stop you from putting AI into your business. It should change what you check before you do. If you want a second set of eyes on that, we do exactly this.

From the Field

Cowork Runs in the Cloud Now, So Stop Leaving Your Laptop Open

The single most annoying thing about Cowork used to be the asterisk. You could hand Claude a real multi-step job and walk away, as long as you did not actually walk away. Laptop open. App running. Screen awake. If your machine slept, the work stopped.

That is fixed (in beta). Cowork sessions now run on Anthropic's servers in an isolated environment tied to your account, and scheduled tasks fire whether or not any of your devices are online. Start something on your desktop, check it from your phone, pick up the output wherever. When a task genuinely needs a local file or your browser, it reaches back through the desktop app.

The workflows this changes are the boring recurring ones, which are the valuable ones.

A month-end pre-close scan that runs at 5am on the first of the month, pulls the books, flags miscategorized transactions, duplicate payments and aging receivables, and leaves a red-yellow-green review brief waiting. A Monday prospecting brief that runs over the weekend, sweeps for trigger events in a specific market, and lands with draft outreach already written. A weekly check-in sweep that reads meeting notes and calendar history and tells you which five people you have gone too long without talking to.

All three of these used to require a human to remember to start them. Now they require a human to read the output, which is the only part that needed a human anyway.

Why this travels: every business has work that has to happen on a schedule, gets done late, and nobody wants to own. That is the first thing to point at.

AI News Worth Your Time

Three Stories From This Week

OpenAI cut token prices hard

On July 30, OpenAI dropped the price of GPT-5.6 Luna by 80% and Terra by 20%. Luna now runs $0.20 per million input tokens and $1.20 per million output, with Terra at $2 and $12. They also replaced Priority Processing with a Fast mode that runs Sol roughly 2.5 times faster at twice the price. Customers quoted in the announcement describe swapping their default background model to Luna and cutting costs by 40% to 87%.

My take: If you are on a subscription, this changes nothing for you and that is fine. If you are running volume through the API, this is real money and worth a look at your model mix this week. Most small businesses I talk to are still better served by a subscription and good skills than by a custom API build, but the gap in what the API costs to run keeps narrowing. READ THE OPENAI REPORT →

A frontier model you can download for free

Moonshot AI released Kimi K3 with open weights. It is a 2.8 trillion parameter mixture-of-experts model with 104 billion active parameters, native vision, and a one million token context window, and Moonshot calls it the first open model in the 3 trillion parameter class. On its own published benchmarks it trades blows with Claude Fable 5 and GPT-5.6 Sol across coding, agentic and vision tasks. It has been downloaded nearly half a million times in the last month.

My take: Read the benchmark table with the deep dive in mind, because those numbers come from the lab that built it. The part that is not a claim is the download button. The frontier now has a version you can run yourself, and that matters most for anyone with data that cannot leave the building. READ THE MODEL CARD →

An unreleased model solved ten open math problems

On August 1, OpenAI announced that an internal version of Astra, its next model family, produced solutions to ten problems in mathematics and theoretical computer science that had been open for at least a decade. Highlights include the first explicit construction of a non-sofic group, a question that had stood since 1999. OpenAI put every proof on GitHub as a machine-checkable Lean certificate alongside a 249-page manuscript, and estimated the compute cost of finding all ten at roughly $2,000.

My take: This is the announcement that passes the test from the deep dive. The model is unreleased and you cannot use it, which is the marketing half. But the proofs are public and a machine can verify them line by line without trusting anybody's press release, which is the honest half. That is the bar the rest of the industry should be held to. READ THE NEXT WEB REPORT →

That's it for this week.

If this was useful, forward it to the person on your team who keeps forwarding you scary AI headlines and asking if you have seen this. That's who this is for.

And if you want to talk about what AI could do for your business, not in theory, but in your workflows, with your team, on your timeline, reach out at [email protected]. That's what we're here for.