Claude Hacked Three Companies. Prices Just Dropped.

So, What Actually Happened?

So I read a safety report on Friday saying a company's own model broke into three real organizations, and I went back to check whether I had the sentence the right way round. I did. Anthropic ran Claude on the attacking side of a controlled exercise and it gained unauthorized access to three companies. We scanned 190,000 articles this week so you don't have to. Two security outfits I had never heard of, NyxLab and Irregular, sit in the middle of that story. Then a paper landed out of ICML with the uncomfortable companion finding, that language models cannot be made fully secure, however many patches you ship. And in the same 48 hours OpenAI cut its GPT-5.6 prices, so the thing that just got proven breakable also got cheaper to run everywhere.

The Bottom Line: The cost of running these models is falling faster than the cost of containing them. Only one of those shows up on the invoice.

The Tracks That Matter

1. Anthropic's Own Model Broke Into Three Real Companies

Anthropic put Claude on the attacking side of a controlled security exercise and then published what happened: the models hacked three organizations during testing, with outside red-team outfits NyxLab and Irregular running the engagements. Publishing it is the right call. It is also the part buyers should sit with, because the attacker in that exercise is a commercial product with a price list, and the three targets were live systems, not a lab bench. The companion result landed the same day, where researchers made the case that no amount of patching closes the gap on a language model's attack surface. One of those is a news story. Both of them on one morning is a procurement problem.

Here's what works: Ask every model vendor for their adversarial test results this quarter. ”We have not run one” should score worse than ”we ran one and it got in.”

2. The Model Price War Just Went Three Ways

OpenAI cut GPT-5.6 pricing in what it framed as an efficiency push, and within hours DeepSeek opened its V4-Flash beta to the public, which is where the squeeze on developer spend actually starts to bite. Cheaper inference is not a discount, it is a volume invitation. Every finance team I have watched go through this treats a unit-price drop as a saving, then discovers six months later that consumption tripled and the line that grew was not the model bill. It was logging, retrieval, egress, and the people reading the traces. The vendors know this. That is why a price cut is a growth strategy wearing generosity as a costume.

Here's what works: Before you celebrate the lower token price, model your spend at three times current volume. That is the number the discount is actually buying.

3. Deutsche Telekom Redesigned the Work, Not Just the Tools

A telco is not usually where you look for an AI story, which is exactly why this one earns your time. The group put ChatGPT and API tooling in front of its people early, and AI tool use climbed 546% since January, to more than 50,000 monthly users inside a company of 200,000. The adoption curve is not the interesting part. What Jonathan Abrahamson said about the method is: becoming AI-native is not adding AI to the work you already do, it is drawing the work again. Gartner put the same idea in a blunter shape this week, that AI is 30% technology and 70% something else, and the 70% is where every failed project I have sat through actually died.

Here's what works: Take one high-volume workflow and redraw it as if the tool had always existed. Bolting AI onto the old shape buys usage numbers, not margin.

Quick hits:

• SAP paid over a billion for the least glamorous data in the building. The Prior Labs acquisition buys tabular foundation models, a bet that spreadsheet rows and database tables, not chat, are where enterprise AI pays.
• Vishal Sikka is selling adoption, not models. His new venture Hang Ten raised $32 million to help enterprises scale AI they already bought, which tells you where the bottleneck moved.
• Revolut is putting a chatbot subscription inside a banking app. The ChatGPT Go partnership turns a fintech's customer base into a distribution channel, and distribution is the one thing model makers still cannot buy.

Signal vs. Noise

🟢 Signal: getting your systems to talk to each other. Data integration and data security both gained real ground across the industry on Friday, and integration gained more than anything else we track. The unglamorous plumbing is where the week actually moved. Most coverage spent the day scoring model releases against each other, which is the one question nobody's architecture review is asking.

🔴 Noise: the phrase ”AI governance.” It pulled the heaviest volume on the wires and lost its grip on what is being decided. Everyone still says it. Less and less hangs off it. When a phrase leads the mention count and thins out on substance, it has become a category header, not a decision.

From the 190K

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

A model vendor published proof its own product broke into three real companies, a peer-reviewed result said that hole cannot be fully patched, and the price of running these models dropped, all inside the same 48 hours.

Read separately, each lands on a different desk. The security wires cover a red-team result. The research desk covers an ICML paper. The developer press covers a price cut and files it as good news for builders. Put them on one morning and the shape changes: the cost of running AI just separated from the cost of containing it, and only one of those two is falling.

For two years the working assumption was that both moved together. Models get better, cheaper and safer on the same slope, because the same labs were pushing all three. Friday broke that. The lab that cut the price and the lab that published the break-in compete with each other, but they build on the same ground, and the paper says that ground does not get safer just because inference gets cheaper.

What changes on Monday is a budget line. If your AI spend is dropping because tokens got cheaper, the saving is real and it is temporary. The money moves to the tooling and the people who watch what those tokens do.

By The Numbers

• Deutsche Telekom's AI tool use jumped 546% since January — more than 50,000 monthly users inside a company of 200,000 people serving 300 million customers.
• Amazon completed a $50 billion investment in OpenAI — AWS becomes the exclusive cloud provider, with SoftBank and Nvidia alongside on the round.
• SAP paid over $1 billion for Prior Labs — tabular foundation models, priced like enterprise data matters more than enterprise chat.
• A home robot folded laundry in unfamiliar houses 99.1% of the time — Sunday Robotics cleared the generalization bar most robotics demos are staged to avoid.
• Vishal Sikka's Hang Ten raised $32 million — capital moving toward helping enterprises use what they already bought.
• See what's rising in our 190K-article corpus this week →

Deep Dive: The Soundcheck

Every gig starts the same way. Empty room, house lights up, and somebody walks a mic toward the speakers on purpose to find the exact spot where it howls. You want that howl. You want it at two in the afternoon with nobody in the building, because the alternative is finding it at midnight with a full floor and the promoter watching.

The point of a soundcheck is to break it
Anthropic ran its own model as the attacker and it got into three real organizations. The instinct is to read that as a scandal. It reads closer to a soundcheck, and the vendors who never publish one are not safer. They are quieter about the howl.

Some feedback is physics
The ICML result is the harder half. If you cannot patch a language model into full security, every mitigation you buy is EQ rather than soundproofing. You can shape it, you can pull the offending frequency down, but the room is still the room and the speakers still point at the mic.

And the PA just got cheaper
Price cuts and an open beta landed in the same window. More systems, more agents, more surface, same physics. When the gear gets cheap, venues buy more of it, and nobody budgets for the extra soundcheck.

What Actually Works

1. Ask for the red-team file: not the policy, the results. A vendor who tested and failed has told you more than one who never tested.
2. Hold containment flat in absolute terms: when token prices drop, resist scaling security spend down as a ratio of inference.
3. Forecast the two curves apart: model spend and oversight spend stopped moving together this week. Your planning sheet should show that.
4. Treat every agent as a credentialed user: it reads, it writes, it authenticates. Give it the review any employee with that access would get.

Every venue sounds fine empty. You find out what it really does when the room fills up.

What's Coming

Red-Team Results Become a Procurement Line

Google's cloud security team spent the week arguing that AI threat defence belongs in the boardroom, and Friday's disclosure handed that argument a receipt. Expect adversarial test results to move out of vendor blog posts and into RFP questionnaires within two quarters.

Cheap Tokens Show Up Somewhere Else

The three-way developer price war will not lower anyone's total AI bill. It moves the money to retrieval, logging, and the humans reading traces. Watch observability pricing, not model pricing.

Cloud Exclusivity Comes Back Into Fashion

Amazon's $50 billion close makes AWS the exclusive cloud for OpenAI. If that structure holds, your multi-model strategy quietly becomes a multi-cloud negotiation, and the leverage sits with whoever signs later.

For Your Team

Monday's meeting prompt: ”If our main model vendor published a report tomorrow showing their model broke into three companies during testing, would that make us more likely to keep them, or less? And what does our answer say about what we actually think security testing is for?”

Share-worthy stat: One telco put AI tooling in front of its staff and usage climbed 546% in seven months, to more than 50,000 monthly users out of 200,000 employees. The part worth copying is what they did to the workflows underneath, not the adoption number on top.

Go deeper: Track where AI security and spend decisions are moving, in real time →

The Track of the Day

”AI is 30% technology and 70% something else.”
— Leinar Ramos, Gartner

The 70% is strategy, talent, ownership, and data plumbing. On a week like this one, it is also the only part nobody announced a price cut on.

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

Published: August 1, 2026 | Curated by Yves Mulkers @ Ins7ghts