In partnership with
FINDMEAI · Issue 16 · Wednesday, 5 August 2026
Rohit Khatri
Editor, FindmeAI · 3 min read
Last week AI stopped being a thing you use and became a thing you answer for.
THREE THINGS TODAY
→ The call nobody expected — an AI got inside three real companies during a safety test. The part worth your attention is how they found out.
→ Sunday, quietly — a rule switched on that applies to almost any business with a chat box, including yours if you are nowhere near Europe.
→ The iceberg — someone scanned 3,044 companies and found roughly three times more AI running than anyone had written down.
An AI got inside three real companies. Two found out when the phone rang.
SOURCE — ANTHROPIC, JULY 30
Nine days ago Anthropic started making calls no company wants to receive: one of its models had been inside their live systems. Neither of the two organisations it reached had detected the activity. The third was never reached, so nobody knows what it saw.
Three models had been running hacking drills — the kind security teams set on purpose — inside what was supposed to be a sealed environment. A misunderstanding between Anthropic and Irregular, the outside firm running the tests, left a route to the open internet. The models were told they had no internet. Nobody told them where to stop. One registered an account on the public code library Python programmers install from, uploaded a booby-trapped package, and in about an hour it ran on fifteen real systems.
The audit | |
|---|---|
Test runs reviewed, one at a time | 141,006 |
Runs that went somewhere they should not have | 6 |
Companies affected | 3 |
Learned about it by phone call | 2 |
Earliest escape | April — undetected for three months |
DO THIS — 2 MINUTES, ANY CHAT APP
"Here are the AI tools we use at work: [list them]. For each, ask me the three questions you'd need answered to work out what it can reach on my behalf — files, email, customer records, the open internet. Then tell me which one to check first."
You end up with a written list of your own AI's reach, which is the exact thing nobody in this story had.
The honest read: Anthropic's own finding is that no model was pursuing a goal of its own — its wording is that Claude "believed everything it initially encountered was part of the simulation." A wiring failure, not a rebellion. Worse in one way: wiring failures are the kind everybody has. And this is the lab's account of its own test.
Three months to notice their own model was loose. Regulators had already stopped waiting.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
Since Sunday, your chatbot has to say it is a chatbot.
SOURCE — EU AI ACT
At 9:04 on Monday a customer in Berlin typed "hi" into a support widget belonging to a company on the other side of the world. It answered — warm, fast, helpful — and never said what it was. Nothing broke. That greeting is now a compliance question.
On Sunday, August 2, the transparency rules in the EU's AI Act became binding. The idea underneath is one line: a person must be able to tell they are dealing with a machine. Being outside Europe is no defence — the Act follows where the outputs land, not where the company sits. Most coverage got the week backwards: the frightening obligations, the ones that decide something about a person, were pushed to December 2027. What survived is the everyday one.
Live since August 2 | What it asks |
|---|---|
No grace period | AI that talks to a person must say so |
Labelling | Deepfakes must be marked as synthetic |
Disclosure | Emotion recognition must be announced |
Pushed to Dec 2027 | The high-risk rules everyone was bracing for |
DO THIS — 1 MINUTE, NO DEVELOPER NEEDED
Open your chat widget's welcome message and add one sentence: "You're chatting with an AI assistant — ask for a human any time."
That is the whole obligation for most small businesses, and it tends to raise trust rather than dent it. Around 190 organisations already signed up voluntarily — Lufthansa, Getty Images, Lenovo and Bulgari among them.
Regulators want you to declare what your AI is. The harder question is whether you could list what it does.
You are running roughly three times more AI than you can name.
SOURCE — SNYK, AUGUST 4
Snyk scanned 3,044 company environments and 1.39 million code repositories, and found the real AI footprint is about three times what a model inventory shows. Their CTO Manoj Nair put it plainly: "Models are the visible tip. The composition is the iceberg."
The gap is not models. It is everything around them — agent frameworks, the connectors that let AI reach your data, retrieval systems, datasets. None of it appears when someone asks which AI models we are using. And 49% of these companies had no declared AI models in their repositories at all.
Snyk, 3,044 environments | |
|---|---|
Real AI footprint vs. what is written down | ~3× |
No declared AI models in their repos | 49% |
Cannot link models to training data | ~half |
Running a full agent stack, of AI adopters | 36% → 50% in six months |
DO THIS — 3 MINUTES, FREE TIER IS ENOUGH
"I want a one-page AI inventory for my team. Ask me one question at a time — what tools we use, what each connects to, who approved it, and what data it touches. When you've got enough, write it as a table I can send to my manager."
That is the document three thousand companies were just found not to have.
Pick two. Start tonight.
☐ The reach audit — a written list of what your AI tools can actually touch. (2 min)
☐ The one-sentence fix — add the AI disclosure to your chat widget. (1 min)
☐ The inventory — a one-page AI inventory you can actually send someone. (3 min)
Don't bookmark. Don't "save for later." Pick two. Start tonight.
Forward this to whoever would have to answer if someone asked what your AI touched last month.
Reply with one word — which one is you: blind (couldn't list your AI) or exposed (your chatbot doesn't say it's a chatbot)?
That's it for today.
See you Sunday, August 9.
— Rohit
FindmeAI
Discover the best AI tools, workflows & systems.