Shadow AI Is Invisible... Until It Costs You $6.37T

So, What Actually Happened?

So I had three tabs open on Tuesday morning and it took me until the second coffee to notice they were one story. Sophos shipped a shadow-AI discovery tool whose whole premise is that you cannot currently see which AI your own staff are using. Then Google started indexing public Claude artifacts, so conversations your team shared last month are searchable by strangers this month. We scanned 190,000 articles this week so you don't have to. Meanwhile Gartner told the market to budget $6.37 trillion for IT in 2026, and Palantir's Alex Karp warned Washington off the European model of regulation. Four stories, one hole in the middle of them.

The Bottom Line: The spend is documented to the decimal. The usage is not documented at all. That gap, not your model choice, is what shows up in the next incident report.

The Tracks That Matter

1. Sophos Ships a Product for the AI Nobody Approved

Sophos launched AI Defense on a premise that reads more like an admission: staff are adopting AI tools faster than anyone can track them, and source code, customer records and financial data are flowing into apps nobody approved. The product finds the activity, weighs it by identity and data access, then enforces policy. What makes it more than a launch is that the category has no settled meaning, one comparison of the tools notes vendors keep stretching the label to fit whatever they already sell. So the first mainstream answer to shadow AI arrives before the market agreed on the question. That ordering matters more than the feature list: whoever writes the discovery report first also gets to define the risk.

Here's what works: Before buying anything, pull last quarter's card statements and outbound DNS logs, and list every AI tool actually in use. That list is your real attack surface.

2. Your Team's Shared AI Chats Just Became Searchable

Google began indexing public Claude artifacts, the shareable pages people generate when handing a colleague a link instead of a file. Nobody was breached. Somebody clicked share, the page was public by design, and a crawler did its job. That is the uncomfortable part, the leak was an interface default rather than an attack. Security vendors have started arguing that controls belong at the decision point instead of the perimeter, and this is the cleanest illustration yet of why. Your firewall never saw the traffic. A person made a two-second choice about a link. Every AI assistant in your organisation has a share button, and very few of them log who pressed it.

Here's what works: Search your company name and project codenames against public AI-share pages this week. Anything that resolves is public, indexed and effectively permanent.

3. Palantir's Karp Tells Washington Not to Copy Europe

Alex Karp told US policymakers to reject the European approach, and his phrasing left no room to misread him: ”We have a template for what doesn't work. It's called Europe.” He also asked the administration not to ban open models. Both halves land hard this week, because a 2.8-trillion-parameter model just shipped with open weights that anyone with the hardware can self-host. Once frontier-scale capability runs inside your own building, the regulatory fight stops being about market access and becomes about what a competent engineer can stand up on a Thursday afternoon without telling anybody. Karp is arguing about the rulebook. The nearer problem is that most organisations could not detect the model even with a perfect one.

Here's what works: Ask your platform team what happens if someone runs an open-weights model on internal hardware today. If the honest answer is ”we'd find out from the power bill”, start there.

Quick hits:

• Apple took the market-cap crown back from Nvidia. The swap came as investors questioned how much Nvidia is investing directly in its own customers, the first time that circularity moved a share price instead of a think piece.
• Cloudflare open-sourced its privacy proxy CLI. The tool hides where a request came from, which is a useful thing to have shipped the same week the industry learned how visible its exhaust already is.
• Singapore's Ropedia raised $22M for physical-AI data plumbing. The pre-Series A is small money aimed at collecting and labelling the real-world data robots need, which is exactly where the shortage sits.

Signal vs. Noise

🟢 Signal: Who can actually see the AI usage. In two days a mainstream endpoint vendor shipped discovery for unapproved AI, a search engine started indexing shared assistant pages, and the security category meant to cover both still has no agreed definition. Most coverage is busy scoring models against each other, which is the one question none of these buyers asked.

🔴 Noise: ”Data quality” as a heading. It pulled heavy volume across the wires again while losing its grip on the decisions that actually moved this week. I have been in this field a long time and it is still the phrase everyone says and nobody owns. The work is real. The word has stopped doing any.

From the 190K

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

A security vendor shipped a tool to find the AI nobody approved, a search engine started indexing what people shared with an AI assistant, and Gartner told the market to plan for $6.37 trillion of IT spending, all inside the same 48 hours.

Read alone, each goes to a different desk. The security trade press writes Sophos as a product update. The consumer-tech desk writes the indexing story as a privacy stumble. The CIO desk writes Gartner as a budget forecast. Put them on one morning and they describe a single imbalance: spending on this stack is growing 14.2% a year, and the register of what is actually running inside the building is maintained by nobody. We have never been this precise about the invoice and this vague about the usage. Every governance framework written in the last two years assumed the reverse, that you would know what you had and the argument would be about permission.

What changes on Thursday is small and slightly embarrassing. Ask your team to name every AI tool touching company data, then compare that list against what expenses and DNS actually show. The distance between the two lists is where your governance programme really starts, and almost nobody has measured it.

By The Numbers

• Worldwide IT spending will grow 14.2% in 2026 to $6.37 trillion — the budget question is settled, the inventory question has not been asked.
• A 2.8-trillion-parameter model shipped with open weights for self-hosting — frontier scale is now something a determined internal team can run on its own hardware.
• Nike was ordered to pay $7.5 million after a jury found gender pay and promotion bias — a reminder that decisions made from company data get litigated, whoever or whatever made them.
• Singapore's Ropedia raised $22 million pre-Series A for physical-AI data infrastructure — capital moving to the unglamorous collection layer that robotics depends on.
• A single commodity desk can have 20 to 50 traders chasing the same phone-call intel — a clean picture of what ungoverned information costs before AI enters the room at all.
• See what's rising in our 190K-article corpus this week →

Deep Dive: Mixing Without a Monitor

Every booth has a monitor speaker. It is not for the crowd, it is for you. It tells you what the room is actually hearing, as opposed to what you believe you are sending. Play one venue where the monitor is dead and you find out fast how much of your confidence came from being able to hear yourself.

The room got louder than the booth
Sophos built a product around the fact that staff adopt AI faster than anyone can track it. That is not a tooling gap, it is a hearing gap. The policy document describes the set you intended to play. In most organisations, nobody can hear the one actually going out.

The leak was a share button
No breach was required for shared assistant pages to land in a search index. Somebody pressed share because it was quicker than attaching a file. The most expensive AI decisions of this year will be made by people who did not know they were making a decision.

Rules do not replace ears
Karp's fight with Europe is about how tight the rulebook should be, and it is a fair argument to have. But a rulebook only governs what you can hear about. Open weights, unapproved apps and one-click sharing all operate below that line, which is why compliance and visibility keep getting mistaken for each other.

What Actually Works

1. Inventory before policy: expenses plus outbound DNS gives you a real tool list in an afternoon. Write policy against that, not against the tools you wish were in use.
2. Audit what is already public: search your own names against public AI-share pages. Treat every hit as published, because it is.
3. Make sharing attributable: pressing share on an AI conversation should be logged and owned, exactly like emailing a file outside the company.
4. Split recommend from execute: water utilities run on one rule, ”AI recommends, humans decide, always”. Copy it wherever money, safety or regulated communication is involved.

You cannot fix a mix you cannot hear. Buy the monitor before you argue about the setlist.

What's Coming

Shadow AI Becomes a Renewal Line Item

Once a mainstream vendor ships discovery for unapproved AI, it stops being a differentiator and turns into a question on everybody else's security questionnaire. Expect ”how do you inventory AI usage” on Q4 renewals, and expect a lot of vendors to answer it badly.

Sharing Defaults Get Quietly Rewritten

The first time someone runs a competitor's name against indexed assistant pages and finds something material, every vendor with a share button changes what ”public” means by default. That is a product change, and it will move faster than any regulation on the subject.

The Open-Weights Argument Moves to Procurement

With self-hostable frontier-scale models in general circulation, ”we cannot run that here” stops being true. The debate shifts from what is legal to who signs off internally, and your platform team has not been invited to that conversation yet.

For Your Team

Thursday's meeting prompt: ”If a journalist searched our company name against public AI-share pages tomorrow, what would they find, and who in this room would already have known about it?”

Share-worthy stat: Gartner expects worldwide IT spending to reach $6.37 trillion in 2026, up 14.2%. Almost nobody spending that money can produce a list of which AI tools are touching their data today. Precise to the decimal on the invoice, blank on the usage.

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

The Track of the Day

”AI recommends. Humans decide. Always.”
— the governance rule water utility operators run on

The clearest sentence about AI accountability this week came from the people who move drinking water, not from a frontier lab. Worth asking why the industry with the least hype was the one that bothered to write the rule down.

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

Published: July 29, 2026 | Curated by Yves Mulkers @ Ins7ghts