Claude now signs everything it writes
Anthropic has started marking Claude's output with an invisible watermark, worldwide, because of a European law. Nobody asked British firms whether they wanted it, and that is increasingly how rules reach them.
- 01
Anthropic starts marking everything Claude writes
Anthropic's documentation, picked up on 11 August, sets out how Claude output gets labelled: an imperceptible watermark woven into generated text, and signed C2PA provenance metadata on generated image files. Models launched on or after 2 August 2026 support it from day one, and it applies worldwide across the API, Claude, Claude Code and cloud partners including AWS and Google Cloud, not only inside the EU. Older models are still being retrofitted with no date given, the detection tooling is promised rather than shipped, and Anthropic is explicit that a mark only means the text may have passed through Claude.
Anthropic - 02
OpenAI ships an offensive security model to vetted users
Announced 10 August. The Daybreak programme splits into two tiers: Blue, which is general models with safeguards tuned for defensive work, and Red, which hands approved users GPT-5.6-Cyber, a model trained to answer the exploit and vulnerability questions other models refuse. OpenAI reports it completed 95% of advanced security tasks against 1.5% for the standard model, says it has already found unreported high-severity bugs in real software including Chrome, and is making hardware security keys mandatory on every Daybreak account from 1 September. Entry needs identity verification, monitoring and a legal declaration, so it is not a public tool, but the capability now exists in a supported product.
OpenAI - 03
North Korean spies were running the models on their own servers
South Korean security firm Genians published findings on 10 August from a spear phishing campaign it calls Operation GitPower, and found local language model setups on the attackers' own infrastructure: Ollama, GPT4All and Msty, plus retrieval tooling pointed at stolen documents and Whisper for turning recordings into searchable text. Running models locally means no provider can cut them off and no usage shows up in anyone's abuse logs. Reuters picked it up a day later and noted the findings are not independently verified, but the practical point holds: bad grammar and odd formatting are finished as a way of spotting a fake email, and what is left is behaviour, like whether the attachment quietly starts PowerShell.
Genians - 04
Meta puts a 30B model under Apache 2.0
Published 10 August: Muse Glimmer is a 30-billion-parameter open-weight model built for agent work, tool calling, coding and multimodal reasoning, trained on data from over 100 languages and released under Apache 2.0. That is a more permissive licence than Llama ever carried. Quantised it needs roughly 20GB, so it runs on a single consumer GPU or a well specified laptop and downloads from Hugging Face today. The point for a small firm is not the benchmark, it is that a model running on your own machine never sends the document anywhere.
Meta AI Research - 05
A ChatGPT Business seat at five times the price, on a waitlist
OpenAI opened a waitlist on 10 August for Premium seats on ChatGPT Business at 125 dollars per user per month, or 100 billed annually, against 25 dollars for a standard seat. What you get for it is five times the usage and no five-hour cap, aimed at people who hit the ceiling most days rather than most months. Read the status carefully: this is a waitlist with early access for some workspaces and no general availability date, and the joining credit offer closes on 20 August.
OpenAI
The watermark is the small part of that story. How it reached you is the part worth ten seconds.
Anthropic did not add this because a customer in Manchester asked for it. Article 50 of the EU AI Act is what made it happen, and rather than run one Claude for Europe and another for everyone else, Anthropic switched the marking on everywhere. So a rule written in Brussels, for a market you may not sell into, lands inside the quote you drafted on Tuesday, through a supplier, while you are in scope of nothing at all. That is now the ordinary route regulation takes into a firm of twelve. Not a letter you have to read. A change in a tool you already pay for. On site the equivalent is a product spec changing because the manufacturer reformulated it, not because anyone on your job asked, and you find out when the delivery turns up looking different.
Worth being plain about what the mark actually is, because the panicked version is already circulating. It says the text may have passed through Claude. It does not say the work is poor, that nobody checked it, or that a client can prove a thing, and Anthropic has published the commitment before the detector. Heavy editing very likely breaks it, which cuts both ways: anyone treating a detection result as proof will accuse the innocent and clear the guilty at roughly the same rate. So the job here is not technical. If a main contractor asks whether AI drafted your method statement or your tender answer, have a plain answer ready, and pick one you are comfortable giving before somebody asks rather than after.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.