AIHOY, AInauts,

Welcome to the latest issue of your favorite newsletter.

Today, we are starting companies, producing audiobooks, and locking down AI assistants. We tested two of the three topics ourselves. For the third, we are honestly glad that security researchers did the testing for us.

Here is what we have for you today:

Let's dive in.

You know the fix: longer prompts, more context, clearer constraints. But typing all that takes five minutes per prompt, so you shortcut it. Every time.

Wispr Flow lets you speak your prompts instead of typing them. Talk through your thinking naturally — include context, constraints, examples — and get clean text ready to paste. No filler words. No cleanup.

Works inside ChatGPT, Claude, Cursor, Windsurf, and every other AI tool. System-level, so there's nothing to install per app. Tap and talk.

Millions of users worldwide. Teams at OpenAI, Vercel, and Clay use Flow daily. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

Let's begin with a practical example of how AI and smart companies are trying to reduce Germany's notorious bureaucracy.

If you have ever started a company in Germany, you know the routine: a notary appointment, forms nobody understands, the commercial register, waiting, a bank account, the Chamber of Commerce, more waiting. Eight weeks or longer is not unusual.

Company formation is only one of many areas where bureaucracy makes everything harder. Several startups are now trying to streamline the process.

One of them is beglaubigt.de, a service designed to make German company formation possible online.

Language note: beglaubigt.de and its checkout are designed for company formation in Germany. The service and some screenshots in this section are in German.

This week's news: beglaubigt.de is now available as an app directly inside ChatGPT.

Tell the assistant that you want to start a company, and the workflow begins.

Naturally, we tested it immediately.

The app guides you through the process step by step: GmbH or UG, company name, business purpose, share capital, shareholders, and managing directors. Everything happens in the chat.

The final checkout opens in a secure browser window. Payment and identity verification happen outside the chat.

The assistant never processes payment details, and a notary is still required. The app coordinates the process instead of pretending the legal steps do not exist.

The company behind the service is Munich-based Openlaw GmbH. According to Y Combinator, it has raised a €3.3 million seed round and works with a network of hundreds of notaries. The promise is to reduce a process that can take eight weeks to roughly three days.

You can install the app from the Plugins area in ChatGPT.

The interesting part is not only that you can start a GmbH from a chat window. The bigger shift is that a complete business workflow begins exactly where the customer expresses intent.

People already ask ChatGPT how to start a company. Now the answer can change from a general guide to: Sure, let's begin.

That matters for every business looking for customers. The service can capture demand before the user opens Google, compares prices, or visits a competitor.

What do potential customers ask an AI assistant before they discover your company? And where could your business move from explaining a process to actually starting it?

We hope this is the beginning of a broader trend: less bureaucracy through well-designed AI workflows.

Claude Code, Codex, and Cursor are on every engineer's stack. Most still treat them like a search bar. Top engineers work from a system, these 100+ prompts are that system. Sign up for The Code and get the prompts free, plus a 5-minute daily newsletter to keep sharpening your edge.

Claim your free prompts

Can you write books with AI? Yes. But can you turn one into an audiobook that people can actually listen to?

We tested the current models and built a complete German-language audiobook: 8 hours and 27 minutes, 20 chapters, mastered and ready for distribution.

The source was one of our unpublished manuscripts with roughly 68,000 words.

The decisive questions were not whether AI could read the text, but which model sounded authentic, how to verify more than eight hours of audio, and what the full production would cost.

Visual note: some filenames and verifier output in the screenshots are in German.

Two lessons from the experiment surprised us.

We ran a blind test in two rounds with eleven candidates from ElevenLabs, MiniMax, and Inworld. Every candidate read the same chapter, and the files were labeled only A through K.

We selected the same voice several times without knowing which provider or model was behind it.

The surprising part: the same voice sounded obviously synthetic with ElevenLabs Multilingual v2, but dramatically more natural with Eleven v3.

The leap in quality came from the model, not the voice.

In our test, ElevenLabs Eleven v3 delivered the best result.

Nobody listens to an 8.5-hour production from beginning to end just to catch every pronunciation error.

So our verifier transcribed every finished chapter again and compared it word for word with the manuscript.

For this second pass, we used the ElevenLabs Scribe speech-to-text API.

The verifier found around 50 cases where the model had read the written form 'above/below' as 'above slash below.'

After correcting those cases, all 20 chapters passed. The word error rate ranged from 0 to 3.4 percent.

The lesson: use AI to check AI-generated work, especially when the output is too long to review manually.

Audiobook production becomes surprisingly straightforward once you know which tools and models to combine.

Our stack:

A strong manuscript can become a complete audiobook in an afternoon. A weak manuscript can become a weak audiobook just as quickly.

One caveat: small artifacts can remain, and as of August 2026, Audible did not accept fully AI-narrated audiobooks. Check the current rules of each distributor before submitting.

Still, the experiment shows how much time and production cost this workflow can save.

We are closing with a security topic that has appeared before, but it is becoming much more relevant as AI assistants gain access to inboxes and workspaces.

Recent coverage and live demonstrations at DEF CON showed how cheaply attackers can search for weaknesses and how email can become an instruction channel for connected AI agents.

The crucial point is this: your AI assistant can become a new front door into your systems.

The EchoLeak vulnerability demonstrated the pattern. A malicious message could influence an AI system without the user clicking anything.

Microsoft patched that specific issue, but the underlying risk remains wherever an AI system reads untrusted content and can also access private data or external tools.

Researchers have demonstrated related attacks through email, Slack messages, support tickets, and even product descriptions.

Models are getting better at recognizing prompt injection, but recognition alone is not a reliable security boundary.

This matters now because many people grant Claude Cowork, ChatGPT Work, and similar tools direct access to email and company data.

Visual note: the permission controls in this screenshot use a German interface.

You could disconnect email entirely, but that removes much of the convenience.

A more practical framework is Simon Willison's lethal trifecta. An AI system becomes especially dangerous when three capabilities are unrestricted at the same time:

Any two can often be managed. All three together create the dangerous combination.

A safer setup looks like this:

This is not an argument against AI employees. It is an argument for giving them the same permission boundaries you would apply to any other system.

The extra confirmation click before an AI assistant sends an email or moves data is a safety brake. Do not disable it simply because it feels inconvenient.

Smarter models are not enough. An attacker can try repeatedly. Your architecture, permissions, and confirmation gates provide the real protection.

That's it for today. Thanks for reading.

See you in the next issue.

Reto & Fabian from AInauten