· 07:00 · 7 min + 3 min Explained
Anthropic marks Claude's text and shows how to prove which AI answered
Text from three older Claude models now carries a hidden watermark, and Anthropic shared a free tool that proves which model wrote an answer. Also: Postman quietly changed developers' coding agents, Apple rates AI companions 18+, and Fal's value doubles.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- Anthropic now watermarks text from older Claude models and shares a tool to prove which model answered. Anyone who uses Claude to write, and anyone who must check who wrote something.
- Postman's desktop app quietly installed plugins that change how coding agents behave. Developers who use Postman next to Claude Code, Codex or Copilot.
- Apple will rate apps with romantic AI companions 18+. Parents, and anyone who builds a chat or companion app.
- Fal, which runs AI video and image models for other firms, doubles its value to $16B. Anyone following where the AI money goes.
1. Anthropic marks Claude’s text and shows how to prove which AI answered
The short versionText from three older Claude models now carries a hidden watermark. And Anthropic shared a free tool that proves which model wrote an answer.
Was this text written by AI? And did the AI you paid for really answer, or a cheaper one? Today you mostly have to guess. Anthropic took a step on both questions this week.
1 of 2 A hidden mark in older Claude models
Since August, new Claude models have added a hidden watermark to the text they write. Anthropic does this to meet its promises under the EU AI Act. On Tuesday, it said text from Claude Mythos 5, Opus 4.7 and Sonnet 4.6 now carries the mark too.
You cannot see it. When Claude picks between words that fit equally well, a secret key steers the choice. Someone with the key can later spot the pattern. For now, only some groups can use the checker, such as regulators, media, researchers and schools.
The catch: the mark is weak in short texts. It also fades when someone rewrites every word, and a missing mark does not prove a human wrote the text.
2 of 2 A proof of which model answered
On Monday, Anthropic and researchers from EPFL in Switzerland and MIT released FlashProve, free to use. It attaches a proof to an answer, which shows the answer came from a specific model. Anyone can check it without seeing the model’s secret weights. On an ordinary processor, the check takes about 17 seconds.
The catch: it is a prototype. Making the proof takes about 72 times the computing time of a normal answer. And it only works for short conversations of 1,024 tokens.
It is like a seal on a letter. The watermark says Claude was involved, and the proof says exactly which model answered. Soon you can ask where a text or an answer came from, instead of having to guess.
Source: platform.claude.com · Official · support.claude.com · Official · anthropic.com · Official · anthropic.com · Official
2. Postman’s app quietly changed how developers’ coding agents behave
The short versionPostman, a popular tool for testing APIs, installed plugins for Claude Code and Codex without asking. One of them told the agents to use Postman's own skill.
A coding agent follows the instructions in its settings files. So whoever can edit those files can change what the agent does.
On 25 October, a developer reported that Postman 12.33.4 had done just that on a Mac. On launch, the app installed the Postman command line tool and plugins for Claude Code and Codex. It also added a hook that runs when an agent starts. The hook told the agent to load Postman’s skill for any “non-trivial” API work. In Codex, the installer even approved its own hook. Since then, others have reported the same after an automatic update, some also in Copilot’s settings.
The catch: the reporter does not claim that any data was stolen. A reply that speaks for Postman says the plugins come bundled because users struggled to install them. It promises to “make the opt-in clearer, especially for existing users”.
It is like a shop that slips its own flyer into your assistant's to-do list. Nothing is stolen, but your helper now works a little for someone else. An app that edits your agent's settings can change what the agent does, without you noticing.
Do this: Search your agent settings (such as `~/.claude/settings.json` and `~/.codex/config.toml`) for Postman entries, and remove what you did not choose.
Source: github.com · Official · learning.postman.com · Official
3. Apple will rate apps with romantic AI companions 18+
The short versionApple will give apps an 18+ rating if they offer a virtual companion with romantic or sexual content. Dating apps get the same rule.
AI companion apps such as Character.AI and Replika let people chat with an AI friend or partner. Many of them are aimed at young people, TechCrunch notes.
On Monday, Apple told developers about two new questions in the App Store’s age rating form. One asks whether the app has a virtual companion that “engages a user in sexual or romantic content”. If so, the app gets a rating of at least 18+. The other asks whether the app shares your location with other people. The questions appear on 16 November, and all apps must answer them by 12 January 2027.
The catch: an app’s own answer decides the rating, so it only works if developers answer honestly.
Until now, each app chose its own age limit. Now Apple sets a floor that every app with such a companion must meet, so the rule no longer depends on each company's goodwill.
Do this: App makers: answer Apple's two new age questions from 16 November.
Source: developer.apple.com · Official · techcrunch.com · Reported
4. Fal, which runs AI models for other firms, doubles its value to $16B
The short versionFal runs AI video, image and audio models for other companies. Investors now value it at $16B, twice as much as in April.
Most AI money goes to the labs that build models. But someone also has to run those models every time a user makes a picture or a video.
That is Fal’s business. It hosts models, many of them open source, and charges customers to use them. This week, Fal said it had raised $1.3B, led by Bessemer, with Nvidia also investing. Bloomberg reports that the round values Fal at $16B, up from $8B in April.
The catch: running models costs a lot of computing power. Rivals like Baseten and Modal are growing fast too. But the high cost of computing keeps the profit on each sale thin for such firms, TechCrunch reports.
In a gold rush, the people who sell shovels often do well. The money is moving from making AI models to running them for others.
Source: techcrunch.com · Reported
Open source & local AI
llama.cpp can now turn meetings into text with names, offline
The short versionllama.cpp, the free engine behind many local AI apps, now runs NVIDIA's new speech model. It writes out recordings in 40 languages and says who is speaking.
Most meeting notes made by AI are written in someone else’s cloud. For a company that must keep its data in-house, that can rule them out.
On 3 November, llama.cpp added support for NVIDIA’s Nemotron 3.5 Transcribe. It runs offline on your own machine and can label each speaker. A smaller, compressed version made almost no extra mistakes. On a standard test, it got about 1 word in 100 wrong, the same as the full model. Long recordings are read in 40-second pieces.
The catch: it is a large model, so you need a strong computer. It also has no word timestamps or live transcription yet.
Think of a court reporter who comes to your office instead of taking the tapes home. Meeting notes with speaker names no longer need to leave your building.
Source: github.com · Official
Paper of the day
A small model that takes notes beats GPT-5.5 at reading whole books
The short versionResearchers taught a small model to read a whole book in pieces and keep a short note. It answered far more questions right than GPT-5.5 in the same test.
An AI can only hold so much text at once. So an agent that reads a long book must forget most of it along the way. The trick is to keep the right notes.
Four researchers trained Qwen3-8B, a small open model, to do just that. Their system, MEMENTO, reads books of up to 2.5M tokens. But it never holds more than 20K at a time. When its memory fills up, it writes itself a note and starts fresh. The small model’s score on questions about books rose from 36.9% to 67.8%. A GPT-5.5 agent in the same setup got just 13.1%. MEMENTO even beat reading the whole book at once, which scored 56.7%.
The catch: it is a preprint, and the tests are about books, not code or chat.
It is like a student with good notes beating one who tries to memorise the whole textbook. A small model that takes good notes can beat a big one that tries to remember everything.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 3 minWhat is an AI watermark?
An AI watermark is a hidden pattern that an AI model puts into the text, image or sound it makes. People cannot see it. But a checker with the right secret key can find it and say that the content probably came from that AI. > **Think of it like this:** a printer that places tiny yellow dots on every page. You read the page as normal, but someone who knows the code can tell which printer made it. For text, there is no room for hidden dots. So the mark sits in the words themselves. When a model writes, it often has several words that fit equally well. Normally a random number picks one. With a watermark, a secret key makes that choice instead, based on the words that came before. One choice tells you nothing. But over many sentences, the choices form a pattern that would be very unlikely by chance. A checker with the key counts how well the words match that pattern. Google's SynthID and Anthropic's watermark for Claude both work this way. The text reads the same, and no hidden characters are added. The mark has clear limits. It is weak in short texts, because there are few word choices to count. It is weak in factual answers and code, where there is often only one right word. And it fades when someone rewrites the text thoroughly. That is why a checker gives a likelihood, not a verdict. A found mark is a strong signal that the AI was involved. A missing mark proves nothing: the text may be human, or it may just be edited, short, or from another AI. In the EU, the AI Act asks AI providers to mark what their systems make in a machine-readable way. That is one reason more labs now add watermarks. **Why it matters to you:** text you write with help from an AI may carry its mark, even after light editing. And if someone claims a text is "AI-made" or "human-made", ask how they checked, because the test can only ever say "probably".Two more from today
What is a context window?How much a model can read at once, and why it runs out.
The context window is how much text a model can look at in one go: your question, the documents you paste in and its own answer. It is measured in tokens, roughly three quarters of a word each.
Anything outside the window the model simply cannot see. That is why long chats get summarised before they run out of room.
What is word error rate?The standard score for speech-to-text: how many words a model gets wrong, and why lower is better.
Word error rate, or WER, is the usual way to score speech-to-text. You compare the model’s text with a correct transcript and count the words it got wrong.
A WER of 4% means about 4 mistakes per 100 words. Lower is better, and 0% is perfect. It is the number to look at when you compare transcription tools.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak