Hello AI Enthusiast,
Two moves from Anthropic this week, pulling in opposite directions. Every new Claude model now leaves an invisible mark on the text it writes, which is the EU transparency rules finally landing inside a product. At the same time, Claude Code is moving inside company firewalls and switching on a mode that stops asking permission. One is about proving where content came from, the other about letting AI work with less supervision. Let's get into it.
The Big Picture 🔊
Claude Now Watermarks Every Word It Writes
Claude now marks the text it writes. Models launched from August 2 embed the mark in the text itself, invisible when you read it and built to survive copy and paste, while image files carry standard C2PA provenance data. It applies worldwide, across Claude, Claude Code, Cowork and the API. Anthropic signed the EU AI Act transparency code, and the tool to detect the mark has not shipped yet.
|
||
|
Labels are already part of how we read things. You see a small "made with AI" note under an ad or a product photo and you keep scrolling. A watermark sits one layer deeper, and it raises a question worth sitting with: once you can tell, what do you do with that? Knowing a text came out of Claude says nothing about whether it is good. Plenty of careful work starts there and gets hours of human editing after. It changes things when something goes wrong. If a text damages someone, or a document turns out to be false, proving how it was made changes the conversation. |
Watermarks put a label on the output, but they say nothing about whether the work behind it was any good. That judgement still sits with your team. Our Corporate AI Training is built around exactly that, knowing which AI output to trust and which to send back.
Claude Code Moves In-House and Stops Asking Permission
Anthropic pushed four Claude Code updates in a week. Companies can run sessions on their own infrastructure instead of Anthropic's, and the Compliance API now records Cowork and Claude Code sessions too. Sessions can hand a summary to each other, so a second one picks up mid task. From August 14, auto mode is the default on Pro, Max and Team, letting Claude work without stopping for approval.
|
||
|
Self hosting and the Compliance API are answers to the two questions that stop regulated companies from saying yes: where does our data live, and can we see how our people use this. The Compliance API is not new. What is new is that it reaches Cowork and Claude Code, which is exactly where those companies were most nervous. Auto mode is the one that surprised us. We assumed fewer interruptions would mean lower cost, and it turns out to be the other way around, because every tool call passes through a safety classifier. Anthropic is absorbing that cost for Pro, Max and Team, which tells you how badly they want people to switch. The testing numbers explain the push. Human review caught 13.6% of dangerous commands. Auto mode blocked 89%. We saw a small version of this ourselves, when Claude flagged that one change would break a scheduled automation we had forgotten about. |
Bits and Bobs 🗞️
Google's Gemini app has passed 1 billion monthly active users, the same milestone ChatGPT reached in June.
OpenAI has improved GPT-5.6 Sol for Plus and Pro users, adding a slider to control how deeply it thinks, and opened GPT-5.6 Luna to free users with unlimited text chats.
ChatGPT Business is getting premium seats with roughly five times the usage of standard seats, something Claude has offered for a while.
OpenAI has released GPT-5.6-Cyber, a model for trusted security teams that helps find vulnerabilities and refuses fewer requests than the standard models.
Meta has open-sourced Muse Glimmer, a 30 billion parameter agent model that runs offline on normal consumer hardware.
xAI has launched Imagine Image 2.0, with focused edits, multi-image references and ready-made templates.
Spotify will label "AI Persona" profiles and keep their music out of editorial and algorithmic recommendations unless you follow them.
Claude helped improve a bound on the Riemann hypothesis from 41.6% to 67.2%, on a problem mathematicians have worked on since 1859.
Demis Hassabis is stepping aside from leading Google DeepMind.
Claude Enterprise now has inference hooks, which let a company's own security server approve or block prompts before the model sees them.
That's a wrap on our newsletter!
If you're looking to build AI skills in your team, check out our Corporate Training. Every program is tailored to your team and built around hands-on practice.
Catch you next week! 👋
