More and more software is written by AI and read by nobody. For a quiz on a news site, that is fine. For a bank account or a payment system, it is not.

In this issue: where vibe coding makes sense, and why giving admin rights to AI is plainly dangerous.

And some good links evaluating the Anthropic IPO, with relevant numbers.

Just as a sign of the times: in October, large US banks found AI agents inside their customers' accounts. The customers had handed over their passwords. Some agents then did things nobody had asked them to do.

AI companies pay for English content, but all other markets and languages are of less importance, at least for now. Of 230 publicly known deals between media and generative AI companies (2023 to May 2026), 71% went to outlets in the US, UK, Canada, Australia and New Zealand. The US alone has 128, India 2. Reporters Without Borders found no public OpenAI deal with an Asian outlet, and no data at all for Africa (RSF).

October 2026 · Axios / Forbes / Sierra / Meta

Major US banks are finding third-party AI agents inside customer accounts, often because customers handed over their login and password. Some agents acted beyond what the user had asked for (Axios, US News). The banks now restrict or block them.

Retail is split. Amazon blocks outside shopping agents completely, QVC lets them buy, Gap decides case by case. Part of it is fraud and liability, part is money: Amazon earns billions from ads, and an agent does not see them (Forbes).

Meta answers with a standard. On 6 October Meta and Sierra announced the "Personal Agent Protocol": how an agent logs in and what it may do (Sierra). Partners include Shopify, Stripe and Walmart, though Shopify and Stripe also joined a rival protocol (The Next Web).

The agents arrive first, the rules come later.

22 October 2026 · ARD, ZDF, Deutschlandradio / Medientage München

Two media economists, Christian Zabel (TH Köln, Cologne's University of Applied Sciences) and Ulrike Rohn (Tallinn University), presented "Der deutsche Medienmarkt 2035" (The German Media Market 2035) at the Medientage in Munich. The analysis presents four different scenarios, built from interviews on 51 trends with RTL, Netflix, Bavaria, the publishers' association BDZV and others (report site). In all four, the market shrinks. In the worst, "Big Tech Takeover", AI becomes the main way people reach information. The number of professional journalists halves, from about 40,000 to 20,000, and revenues fall 43%. In the best scenario, "Level Playing Field", revenues fall about 10% and about 35,000 journalists remain (dpa via finanzen.net). Competition lawyer Thomas Höppner puts the difference between the worst and the best case at EUR 8.5 billion in revenue a year: money that either stays with German media or goes to the platforms (LinkedIn).

One possible path: agreements on pay-per-use and pay-per-crawl, transparent recommender systems, room for media to cooperate. Note: Germany's public broadcasters ARD, ZDF and Deutschlandradio commissioned the report, and it casts public broadcasting as the market's stabiliser. (Disclaimer: my employer Deutsche Welle (DW) is a member of ARD.)

In a controlled trial, experienced developers worked 19% slower with AI tools. Afterwards they estimated AI had made them 20% faster (METR). That was early 2025. In a 2026 rerun METR could no longer measure the effect cleanly and now says AI "likely" brings a gain (METR).

That gap might be the best possible short summary of vibe coding. There is really no question that AI increasingly writes code. But we need to ask: what kind of code? Not all of it is created equal. Is it useful? Yes. Is it safe? That depends on what you build.

Below I tried to differentiate between three different levels, and a contrarian view regarding future code security.

On this level a non-coder such as a journalist builds a quiz, a dashboard, a small checking tool. This is fast, easy to implement, no big risks. Nieman Lab expected this to be the big newsroom shift of 2026 (Nieman Lab).

My experience, recently: for ChatEurope I built a small app to enable humans to review the chatbot answers. It is basically an elegant spreadsheet, but with real functions and a better interface than I could have built before (made with Claude, which I also use for this newsletter).

But go just one step further with the same type of mini-application, and the problems start. The same tool on the site of a known media company must trigger reviews: security, data protection, integration into the content management system (CMS). Most newsrooms have no process for that yet. If 50 vibe-coded apps land on an IT department's desk in a month or even a week, the review takes months, and by then the energy and the benefits are gone.

In short: as long as no personal data, payment data or credit cards are involved, the risk is manageable and this should be ok. But note: reports of data leaks from amateur apps are growing (Tech Times).

Here the key question is who writes the code and who checks it. If a team uses AI to add features and still reviews them, it can be a real gain.

There is a cost to skills. In a study by Anthropic (the company behind Claude), people who coded with AI scored 17% lower on a quiz about the concepts they had just used (Anthropic).

This is where it gets serious. Code goes into complex, existing systems, and nobody reads it line by line.

Veracode tested the code models produce: in about 44% of tasks it contained a known security flaw. Barely better than in the first report (Veracode). The Cloud Security Alliance counts officially registered vulnerabilities traced to AI code: six in January 2026, fifteen in February (CSA).

The people who would catch such flaws are getting fewer. At Anthropic, the share of junior hires in technical roles fell from 27% to 9%. In August, its AI agents logged six times more hours than its staff (Anthropic).

There is a fair counter-argument. The same tools find old holes. FIRST, the global network of security teams, calls 2026 the "AI discovery era". Mozilla saw a 164% jump in reported vulnerabilities in the first quarter (FIRST).

The best example came on 21 October. Two critical holes in PHP, the language behind roughly 70% of all websites, had sat in every version since 2004. The security firm Calif found one with an OpenAI model, Wiz found more with its own AI agent. The US National Security Agency advised: if you cannot patch, take the site offline (TechCrunch). Wiz adds the uncomfortable part: before AI, these holes were too hard to exploit. Now a language model can adapt the attack to each server (Wiz).

The PHP Foundation's own account is more sobering. Its team had seen first signs of the problem in 2024, after an audit funded by Germany's Sovereign Tech Agency. The fix stalled for months, because only a few people understand that part of the engine (PHP Foundation). Since July the team has published 15 CVEs (Common Vulnerabilities and Exposures, the public catalogue of known security flaws, cve.org) and worked through more than 100 reports that were invalid, duplicates or not reproducible. Their summary: tools find bugs faster than ever, but deciding on the right fix takes experienced engineers, and there are too few of them. (Anthropic's Claude scanned PHP earlier this year, too.)

To me it feels like light and shadow at the same time. Small beginnings work, and creativity in small steps is welcome. But whoever sees these first apps as a replacement for professional developers is on the wrong road. I expect it will take two to five years before vibe coding is built into security and review processes.

Georgetown University's Center for Security and Emerging Technology (CSET) on three kinds of risk in AI code (CSET).

Dutch language model: The Netherlands builds its own language model for Dutch, run by the non-profit research organisation TNO, the Dutch universities' IT cooperative SURF and the Netherlands Forensic Institute, with a budget of EUR 13.5 million (TNO, SURF). It trains only on licensed, open or permitted data, no scraping. Dutch news publishers license their archives through the association NDP Nieuwsmedia, which the founders call the biggest milestone so far (IO+). Pilots with government agencies have run since February 2026, a wider roll-out is planned for the second half of the year. Small next to the big models, but one of the few where publishers are paid partners, not raw material.

Pop stars as curators: Dua Lipa curates this year's London Literature Festival, including a showcase for translated fiction (Official Charts).

Big spending, unclear pay-off: Gartner forecasts that AI data centre owners would need $8.5 trillion in revenue through 2030 to cover their spending and financing. It expects AI revenue of at most $3.6 trillion (Gartner). At its IT Symposium, Gartner also reported that 81% of chief information officers (CIOs) are raising AI spend, but only 13% see significant value (via The Deep View). Apollo's chief economist Torsten Slok sees "zero financial payoff" outside tech: revenue per worker rises for the seven largest tech firms, but stays flat for the other 493 in the S&P 500 (Apollo). Forecasts, surveys and rough market data, not measured projects. But it fits the METR gap in the Talk.

Mega IPO: Anthropic has filed for an IPO (initial public offering, its stock market listing), reportedly aiming at a $2 trillion valuation (TechCrunch). What is remarkable: the company has a growth story never seen before. But it also has an investment plan into the future never seen before. Evan Armstrong has a great write-up, trying to make sense of it all (Is Anthropic Worth $2 Trillion?). From the filing (the S-1, the prospectus every US company must publish before listing): $131.7 billion in cash, about $16.8 billion revenue in the third quarter (press-reported, not yet in the filing), but compute contracts (read: costs) that rise to about $120 billion a year next year, mostly non-cancelable (S-1). Models older than six months bring less than 20% of revenue.
It might all work out, still. But there are gigantic risks. The key formula, according to the Financial Times: in early 2025 Anthropic spent about $2.50 on compute and infrastructure for every $1 of revenue, by mid-2026 it was under 50 cents. Profit is possible only if that ratio stays healthy. Disclosure: I use Claude (Anthropic) for this newsletter.

EMNLP 2026 · 24 to 29 October 2026 · Budapest · 2026.emnlp.org · Still running when this goes out. Watch the papers on low-resource languages and agent evaluation.

Languages & The Media · 4 to 6 November 2026 · London · languages-media.com · Disclosure: I have a speaker slot.

Web Summit · 9 to 12 November 2026 · Lisbon · websummit.com · Watch how much of the agent talk survives questions about liability.

tcworld conference · 10 to 12 November 2026 · Stuttgart · tcworldconference.tekom.de · Technical writers and terminology people, the ones who feed structured content into AI.

JournalismAI Festival · 17 to 18 November 2026 · London and online · journalismai.info · Newsroom AI in practice. Supported by Google.

Language Intelligence 2026 · 17 to 19 November 2026 · Vienna · language-intelligence.com · Panel on why language models need structured knowledge.

On 5 October a YouTube channel called Degenerative Pixels posted "LGTM (Looks Good to Me)", a song made with Suno and animated with Claude (YouTube). The song's story: after a seminar, the boss fires all the developers, and soon "nobody knows what the code is inside". The music is not bad, the lyrics are funny. All of it done by AI.

“Let’s ship it to prod” :-)

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany. Three projects you will hear about in this newsletter:

AI use: Full disclosure, I do use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. My goal is to find out where AI is reliable and where the hallucinations come in. Before publishing I check all facts and links.

Error log: To track the issues of using AI I maintain an open Google Doc where I collect the small and big problems that a tool like Claude introduces to editorial work. The idea is to get a better understanding of where AI is good and where it is not. Read it here.

Note: The newsletter is published on a dedicated website and on LinkedIn. Select the one that is easier for you to read.

babylon-newsletter.com · 7,000 languages. AI works for 20.