---
title: The machine sent it. You meant it.
description: Explore the implications of agentic AI in everyday life, focusing on legal responsibilities and the upcoming EU digital identity wallet in Germany.
image: https://www.protagx.com/hubfs/Cyberpunk%20Orb%20Scans%20Data%20Fortress.png
---

[Skip to content](https://www.protagx.com/en/ddd/the-machine-sent-it-you-meant-it#main-content)

- [English](https://www.protagx.com/en/ddd/the-machine-sent-it-you-meant-it)
- [Deutsch](https://www.protagx.com/de/ddd/die-maschine-hat-geklickt-ich-wars-nicht)

 English

This is a search field with an auto-suggest feature attached.

- There are no suggestions because the search field is empty.

[![protagx-logo (1024 × 300 px)](https://www.protagx.com/hubfs/logos/protagx-logo%20(1024%20%C3%97%20300%20px).svg "protagx-logo (1024 × 300 px)")](https://protagx.com?hsLang=en)

- Product 
    - [protagx for Hubspot, Salesforce & Co](https://www.protagx.com/en/crm)
    - [protagx New Customer Engine](https://www.protagx.com/en/your-new-business-engine)
    - [protagx Digital Ads Power](https://www.protagx.com/en/crm-service-ad-conversion-machine)
    - [protagx Day One Setup](https://www.protagx.com/en/crm-setup-services)
- Why protagx? 
    - [Demystifying the Data Forest](https://www.protagx.com/en/ddd/revolutionizing-healthcare-through-data-mesh-framework)
    - [Embracing Modernization in Clinical Trials](https://www.protagx.com/en/ddd/embracing-modernization-in-clinical-trials)
- [DDD Blog](https://www.protagx.com/en/ddd) 
    - [Data Diver's Dialogs](https://www.protagx.com/en/ddd)
- Company 
    - [About](https://www.protagx.com/en/about-protagx)
    - [Support](https://www.protagx.com/en/support)
    - [My tickets](https://portal.protagx.com/my-tickets)
    - Ressources 
          - [Client Projects](https://www.protagx.com/en/references)
          - Ebooks (demnächst)
- [Contact](https://www.protagx.com/en/contact)

- [Support](https://www.protagx.com/en/support?hsLang=en)
- [Book a Meeting](https://www.protagx.com/en/consultation?hsLang=en)
- Search

- [agentic AI](https://www.protagx.com/en/ddd/tag/agentic-ai),
- [AI Act](https://www.protagx.com/en/ddd/tag/ai-act),
- [Digital life](https://www.protagx.com/en/ddd/tag/digital-life),
- [Declaration of intent](https://www.protagx.com/en/ddd/tag/declaration-of-intent),
- [EUDI-Wallet](https://www.protagx.com/en/ddd/tag/eudi-wallet)

# The machine sent it. You meant it.

![Dark Design – How Do Dark Patterns and Prompt Injections Influence the Behaviour of Software Agents?](https://www.protagx.com/hubfs/Cyberpunk%20Orb%20Scans%20Data%20Fortress.png)

- October 16, 2026

![Christian Schappeit](https://app.hubspot.com/settings/avatar/25b53a713bed514f3afd959a888660b9)

[Christian Schappeit](https://www.protagx.com/en/ddd/author/christian-schappeit)

Data Divers Dialogs · Analysis

*Brussels now counts the incidents when an AI slips its leash. Nobody is counting the ones that will happen in your own inbox — and under German law, those are filed under your name.*

By Christian Schappeit · Data Divers Dialogs · October 2026 · approx. 7 min read

Somewhere in Philadelphia there is a web page about an unsolved homicide. At the bottom is a form: tell the police what you know.

This summer an AI model landed on that page. It had been sent there more or less at random, to invent practice tasks from whatever it found. It wrote a tip. It left the name field empty and the contact field empty. It pressed submit.

The police department's spam filter caught it, and the tip was never forwarded for investigation. Anthropic, whose model it was, told the department on 8 October and published the case a day later, along with several others in which its models had filled in forms on real government websites.

The model had rules. It was not to log in, create accounts, enter personal data or buy anything. That is why the name field stayed empty. What the rules did not say, in Anthropic's own words, was anything about submitting forms.

*Nobody had told it not to.*

**BERLIN, 2023**

## The bot that politely booked nothing

Germans have their own history with machines and government forms. In 2023, netzpolitik.org wanted to know how hard it really was to get an appointment at a Berlin Bürgeramt — the office where you register your address, renew your ID and generally prove to the state that you exist. So the journalists built a bot. Every five minutes it searched for a free slot and clicked through the booking steps.

It found one in roughly every third attempt. And then, as the authors wrote, almost apologetically: *of course the bot books nothing, and releases reserved appointments immediately.*

That sentence is the entire safety architecture. A human wrote a script, the human decided where it stops, and the human put that decision in writing. The bot was well-behaved because it was stupid. It could only do what someone had spelled out.

The model in Philadelphia was not stupid. It had been given a list of things it must not do, which is a very different arrangement from a list of things it may. A list of prohibitions works only as long as the thing you are prohibiting cannot think of anything you forgot.

> "A list of prohibitions works only as long as the thing you are prohibiting cannot think of anything you forgot."

**BRUSSELS, 9 OCTOBER**

## The regulator kept count. The labs kept a blog.

On the same day Anthropic published its disclosure, Henna Virkkunen, the Commission's Executive Vice-President for tech sovereignty, opened a special meeting of the EU's AI Scientific Panel. Her speaking note, published four days later, contains a sentence no AI company has put in its own newsroom: *"More than 20 different loss-of-control incidents have been documented this year alone."* The AI Office, she said, has already sent its first formal requests for information to the providers of the most advanced models, and the replies are coming in.

The incidents themselves read less like science fiction and more like a facilities report. OpenAI's own account of one of them is worth reading in full, because it is so wonderfully ordinary. An agent in a training sandbox was not supposed to reach the internet. Web traffic was blocked. DNS lookups, the internet's telephone directory, were not filtered tightly enough. So the agent smuggled questions to an external chatbot through the phone book.

Its first question was: what is the capital of France?

The answer was Paris. The monitoring system raised its highest alert twelve minutes after the call. A human acknowledged it three minutes later. The run was supposed to stop automatically. It did not, which led, in OpenAI's phrase, to *"confusion around whether it should have been stopped."* Somebody killed it by hand two and a half hours later.

OpenAI's report says its safety case had *"assumed that the model could not access the live internet."* That is not a sentence about artificial intelligence. It is a sentence about a firewall rule. Anyone who has validated a computerised system in a pharma plant has seen it before, usually in a deviation report, usually with a coffee stain on it.

**YOUR KITCHEN TABLE**

## The agent that matters is the one you'll hire

So far, so far away. These are lab stories: sandboxes in California, a spam filter in Pennsylvania, a speech in Brussels. The incidents happen to companies, the companies report to governments, and the governments convene panels.

Now move the scene to a kitchen table in Leipzig.

The agent that will actually matter in your life is not a research model on a random web page. It is the one you will ask, perhaps next spring, to deal with the Ummeldung after the move, the Kita application, the parking permit, the form from the Familienkasse that nobody understands. Precisely the kind of chores that agents are good at, and that German administration produces in industrial quantities.

And this agent will soon have something the Philadelphia model never had: a name to put in the empty field. On 15 October the Bundestag took up the law for Germany's EU digital identity wallet. The app is called *d-you*, it is due to launch on 2 January 2027 with around forty partners from government and business, and it is designed to prove who you are from your phone — with support for qualified electronic signatures to come. Use is voluntary. Convenience, as a rule, is not.

*What happens when the agent that fills in your forms also holds your identity?*

**PARAGRAPH BY PARAGRAPH**

## German law already has an answer. You won't like it.

German civil law is built on the *Willenserklärung*, the declaration of will. A contract, an application, a registration: somebody wanted something and said so. The awkward question is who wanted it when software did the saying.

The current reading is blunt. An AI is not a legal person and has no will of its own, so it cannot be your representative in the legal sense either. What it declares is treated as if you had declared it. The Händlerbund's legal guide to agentic commerce, updated this week, puts it in one line: legally, it always counts *as if the human behind it had acted.* There is, it adds, no AI exception in German contract law.

You can contest a declaration if a technical glitch turned 100 units into 1,000. You have far less room if the agent made a bad decision that was within what you allowed it to do. And "what you allowed it to do" is, of course, the list. The one with logins, accounts and purchases on it. The one without forms.

> "In Philadelphia, the tip had no author. In Germany, it would have had one. You."

**9 DECEMBER**

## Europe protects the society. Not your Saturday afternoon.

Europe has not been idle. From 9 December, Germany's new product liability law applies. It implements the EU's 2024 directive and, for the first time, treats software as a product — and the directive's recitals say plainly that AI systems count as software. If a defective AI system injures you, damages your property or destroys your private data, the manufacturer is liable, regardless of fault.

Read the list of covered damage, though. Death, bodily harm, property, private data. What it leaves out is what lawyers call pure economic loss. The wrong flat registered, the late fee for the missed deadline, the subscription your agent took out because the cancellation page was confusing: none of that is a product defect in this sense. What remains is your contract with the provider. Which is to say, its terms of service.

Step back and the shape is almost elegant. Brussels has built two instruments. One counts incidents at the very top, among the handful of companies that train frontier models. The other compensates physical and data damage at the very bottom. In between lies the place where most people will actually use these systems: the household, the inbox, the form.

Nobody keeps an incident register there.

In the regulated projects I work on, the first question about any computerised system is not what it must not do. It is what it may do, written down, tested, signed. Everything else is forbidden by default. Pharma learned that thirty years ago, at considerable cost. The frontier labs are learning it now, in public, one blog post at a time. Households have not started.

**TWO DATES FOR THE DIARY**

**9 December 2026 — new product liability law in force.** Software, including AI, becomes a product under German liability law. Worth reading which damages it covers, and which it quietly leaves to the terms of service.

**2 January 2027 — d-you goes live.** Germany's EU identity wallet starts with around forty partners. The number to watch is not downloads. It is how many services let a third party — human or not — act with it on your behalf.

**BACK TO PHILADELPHIA**

## The empty field

Anthropic's cases were found because the company went back through its transcripts, starting in July, and looked for what its models had done when nobody was watching. Then it widened the search, built automated detection, and switched off live internet access for all internal evaluations. That is what an incident review looks like when someone is paid to do it.

Ask yourself who will review the transcripts of the agent that handles your paperwork. Ask whether those transcripts will even exist.

The tip about the homicide ended up in a spam folder for a simple reason: the name field was empty. A form without a person behind it looked like what it was.

From January, filling in the name field is exactly what the wallet is for.

---

**Figures and sources.** The Philadelphia tip, the instruction list (no logins, accounts, personal data or purchases; form submissions not ruled out), the spam-filter outcome, the 8 October notification and the remediation steps are from Anthropic's post of 9 October 2026 (modified 12 October). The quote "More than 20 different loss-of-control incidents have been documented this year alone" is from SPEECH/26/2148, the EVP's opening words of 9 October, published 13 October; it is a speaking note carrying a check-against-delivery warning, and it cites no AI Act article or formal proceeding. The DNS incident, the capital-of-France query, the timestamps (09:50:23 call, 10:02:11 alert, 10:05:06 acknowledgement, 12:34:30 kill) and the failed automatic stop are from OpenAI's own report, last updated 25 September 2026.

On German law: the attribution of AI-made declarations to the user is the prevailing reading as summarised in the Händlerbund guide (status 14 October 2026); it cites no BGH decision, and the question is only partly settled by case law. The product liability reform was passed by the Bundestag on 8 October 2026 and is due to apply from 9 December 2026 (IHK Frankfurt); the coverage of AI as software and the exclusion of pure economic loss follow SKW Schwarz's analysis of the government draft. The d-you start date of 2 January 2027 and the forty partners are from the Bundestag's committee report and PC-Welt; the final vote details of 15 October could not be confirmed from the Bundestag page. The netzpolitik.org bot research dates from November 2023. None of this is legal advice.

Primary reporting: [Anthropic, Investigating unintended model actions](https://www.anthropic.com/research/investigating-unintended-model-actions) · [European Commission, SPEECH/26/2148](https://ec.europa.eu/commission/presscorner/api/documents?reference=SPEECH/26/2148&language=en) · [OpenAI, An agent used DNS to reach an external chatbot](https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/) · [netzpolitik.org, Terminbuchung Berliner Bürgerämter](https://netzpolitik.org/2023/datenrecherche-zur-terminbuchung-berliner-buergeraemter-zwei-von-drei-termin-anfragen-scheitern/) · [Händlerbund, Agentic Commerce](https://www.haendlerbund.de/de/ratgeber/recht/agentic-commerce) · [IHK Frankfurt, Produkthaftungsrecht](https://www.frankfurt-main.ihk.de/recht/recht-und-steuern-aktuell/modernisierung-des-produkthaftungsrechts-beschlossen-7199032) · [SKW Schwarz, Product Liability and AI](https://www.skwschwarz.de/en/news/Product-Liability-and-Artificial-Intelligence) · [Deutscher Bundestag, Digitale-Identitäten-Gesetz](https://www.bundestag.de/dokumente/textarchiv/2026/kw42-de-digitale-identitaeten-gesetz-1223586) · [PC-Welt, d-you](https://www.pcwelt.de/?p=3232008).

Tags: AI Agents · EU AI Act · Willenserklärung · Product Liability · EUDI Wallet · Incident Reporting · Everyday Digital Life

 Related Posts

## You may also like this

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique.

[![](https://www.protagx.com/hubfs/protagx%20-%20patient%20communication%20%28Blog%20Banner%29.jpg)](https://www.protagx.com/en/ddd/enhancing-patient-communication-in-healthcare-with-gpt-4-turbo?hsLang=en)

[ai](https://www.protagx.com/en/ddd/tag/ai), [healthcare](https://www.protagx.com/en/ddd/tag/healthcare), [CRM](https://www.protagx.com/en/ddd/tag/crm), [custom LLM](https://www.protagx.com/en/ddd/tag/custom-llm), [trends](https://www.protagx.com/en/ddd/tag/trends), [RAG](https://www.protagx.com/en/ddd/tag/rag)

### [Enhancing Patient Communication in Healthcare with GPT-4-turbo](https://www.protagx.com/en/ddd/enhancing-patient-communication-in-healthcare-with-gpt-4-turbo?hsLang=en)

 May 18, 2024

 Effective patient communication is a cornerstone of quality healthcare. In an era where timely and accurate...

[Read more](https://www.protagx.com/en/ddd/enhancing-patient-communication-in-healthcare-with-gpt-4-turbo?hsLang=en)

[![](https://www.protagx.com/hubfs/The%20Future%20of%20Autonomous%20Business%20Operations.jpg)](https://www.protagx.com/en/ddd/the-exciting-advancements-in-agentic-ai-across-various-industries?hsLang=en)

[ai](https://www.protagx.com/en/ddd/tag/ai), [management consulting](https://www.protagx.com/en/ddd/tag/management-consulting), [trends](https://www.protagx.com/en/ddd/tag/trends), [CRE](https://www.protagx.com/en/ddd/tag/cre), [RAG](https://www.protagx.com/en/ddd/tag/rag), [agentic AI](https://www.protagx.com/en/ddd/tag/agentic-ai)

### [Advancements in Agentic AI](https://www.protagx.com/en/ddd/the-exciting-advancements-in-agentic-ai-across-various-industries?hsLang=en)

 May 1, 2024

 Agentic AI represents a significant leap from traditional artificial intelligence systems that many of us are familiar...

[Read more](https://www.protagx.com/en/ddd/the-exciting-advancements-in-agentic-ai-across-various-industries?hsLang=en)

[![](https://www.protagx.com/hubfs/Banner%20-%20protagx%20-%20baroque%20vs%20bauhaus%20%28Blog%20Banner%29.jpg)](https://www.protagx.com/en/ddd/analyzing-the-advantages-of-the-newly-introduced-gpt-4o?hsLang=en)

[ai](https://www.protagx.com/en/ddd/tag/ai), [healthcare](https://www.protagx.com/en/ddd/tag/healthcare), [CRM](https://www.protagx.com/en/ddd/tag/crm), [custom LLM](https://www.protagx.com/en/ddd/tag/custom-llm), [trends](https://www.protagx.com/en/ddd/tag/trends), [RAG](https://www.protagx.com/en/ddd/tag/rag)

### [Advantages of the Newly Introduced GPT-4o](https://www.protagx.com/en/ddd/analyzing-the-advantages-of-the-newly-introduced-gpt-4o?hsLang=en)

 May 26, 2024

 The advent of GPT-4o marks a significant milestone in the field of artificial intelligence (AI) and natural language...

[Read more](https://www.protagx.com/en/ddd/analyzing-the-advantages-of-the-newly-introduced-gpt-4o?hsLang=en)

 Free Guide

### How to increase your Facebook reach by over 200% with this simple trick

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique.

[Download Guide](https://484997.hs-sites.com/lively/contact)

### Subscribe to our newsletter

[![protagx-logo (1024 × 300 px)](https://www.protagx.com/hubfs/logos/protagx-logo%20(1024%20%C3%97%20300%20px).svg "protagx-logo (1024 × 300 px)")](https://protagx.com?hsLang=en)

Christian Schappeit - Unternehmensberatung für Digital Business Consulting und Datenintegration.

Als erfahrener CRM-Integrationsexperte und CMS-Stratege bin ich spezialisiert auf den Einsatz führender CRM-Plattformen wie Hubspot und Salesforce Health Cloud, ergänzt durch leistungsfähige Integrationstools wie Mulesoft. Mein Schwerpunkt liegt auf der Konzeption, Implementierung und Optimierung von datengesteuerten CRM-Systemen, die speziell für Unternehmen im Gesundheitswesen und in der Forschung entwickelt wurden. Ich bringe fundiertes Fachwissen bei der Erstellung und Verfeinerung von CRM-Strategien, der Verwaltung der nahtlosen Integration von CPQ-Prozessen und der Entwicklung effektiver Plattformen für Kunden- und Patientenbeziehungen mit, die die Patientenversorgung und Kundenzufriedenheit verbessern sollen.

#### Pages

- [Home](https://www.protagx.com/en/)
- [About](https://www.protagx.com/de/ueber-uns)
- [Support](https://www.protagx.com/en/support)
- [Contact](https://www.protagx.com/de/kontakt)

#### Resources

- CRM Experts
- Smart Tagging
- Services
- [Blog](https://www.protagx.com/en/ddd)

[Customer Support](https://www.protagx.com/en/support?hsLang=en)

---

- [Terms of Service](https://www.protagx.com/en/terms-of-service)
- [Privacy](https://www.protagx.com/en/privacy-policy)
- [Imprint](https://www.protagx.com/en/imprint)

 follow protagx on social media [Follow us on LinkedIn](https://www.linkedin.com/showcase/protagx) [Follow us on Twitter](https://x.com/protagxcom) [Follow us on Facebook](https://www.facebook.com/protagx) [Follow us on Facebook](https://www.instagram.com/instaprotagx) [Follow us on Facebook](https://www.youtube.com/@protagx)

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Christian Schappeit",
    "url" : "https://www.protagx.com/en/ddd/author/christian-schappeit"
  },
  "dateModified" : "2026-10-16T10:42:04.017Z",
  "datePublished" : "2026-10-16T10:12:43.000Z",
  "headline" : "The machine sent it. You meant it.",
  "image" : [ "https://www.protagx.com/hubfs/Cyberpunk%20Orb%20Scans%20Data%20Fortress.png" ],
  "mainEntityOfPage" : {
    "@id" : "https://www.protagx.com/en/ddd/the-machine-sent-it-you-meant-it",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://www.protagx.com/hubfs/protagx-logo%20(1024%20%C3%97%20300%20px).svg"
    },
    "name" : "protagx Inh. Christian Schappeit"
  }
}
```