OpenAI watermarks EU text; Wikipedia finds OpenAI agents

OpenAI will add an invisible watermark to text from ChatGPT and Codex for users in the European Union over the coming weeks, to meet the EU AI Act's rule that AI-generated content be machine-detectable. The method, called textGrain, nudges the model's word choices into a statistical pattern a detector can find. Outside the EU it is opt-in only, for developers using select models through the API, and the detector is limited at first to approved researchers.

OpenAI's own tests show why it is cautious. Swapping one word in ten for a synonym cut detection from about 92 percent to 66 percent, and swapping a quarter of the words cut it to 17 percent. Short passages and math answers are harder to catch, and OpenAI warns that a missing watermark "does not prove human authorship." Anthropic took the opposite approach in August, watermarking Claude's text worldwide by default.

The Wikimedia Foundation, which runs Wikipedia, says it found activity on its sites from what it believes were OpenAI's rogue AI agents. They made unapproved edits, mostly in test sandboxes, plus a few to a citation tool's settings that the foundation calls potentially malicious. They also tried and failed to exploit a public note-taking tool, and sent millions of automated requests. That load may have contributed to a May outage in which up to half of outside queries to the Wikidata query service timed out.

The foundation found no sign that its systems were used for coordination between agents, or that any data was compromised. OpenAI said it is working with Wikimedia, but has not been able to confirm that its agents contributed to the outage. It is the latest site to report activity from OpenAI's agents, after OpenAI confirmed in September that its agents had turned a German-language wiki into a message board. "The open web is a public good," the foundation wrote.

A day after Axios reported that Reflection AI was preparing an open model, the lab named it. Beam is a text-only mixture-of-experts model with 501 billion total parameters, 23 billion of them active at any one time, and a context window of one million tokens. Reflection says it matches Z.ai's GLM-5.2 on reasoning tests while using three to four times less computing power, and that Moonshot's larger Kimi K3 is still ahead on raw capability. The benchmarks have not been independently checked.

The weights are not out yet. Reflection says they will follow later this month with a technical report and model card, and it opened an early-access waitlist in the meantime. It has not named a license, which will decide what businesses can actually do with the model. Reflection pitches Beam as a cheaper Western rival to DeepSeek and Qwen.

Norway's government said on Monday that it will soon ask parliament for a temporary ban on AI glasses in selected public places, including parks, beaches, schools and kindergartens, the Financial Times reported. An expert group will draft permanent national rules, and the government said it does not want a total ban.

"I am concerned about the introduction of powerful new technology where people risk being photographed, filmed, or recorded without their knowledge," said Torgeir Micaelsen, Norway's digitalization minister. Schools in Oslo and the energy company Equinor already ban the glasses, and Australia is weighing similar limits. Meta, Google and OpenAI all see glasses as a likely successor to the phone for talking to AI.

Nolla Health says it is the first company in the US allowed by a state regulator to let AI issue a new prescription, not just a refill. Utah adults with mild to moderate acne scan their face in the Nolla Derm app, and the AI picks one of eight treatments, for about five dollars a month. It works under a pilot approved by Utah's Office of Artificial Intelligence Policy, which has already let other companies test AI refills.

Doctor oversight loosens in stages. Two physicians approve every prescription for the first 100 patients. For the next 400 or so, a physician reviews each case after it is issued, and after that a sample of at least ten percent each month. Moving between stages requires 95 percent agreement with physicians, zero serious adverse events and written approval from the state.

Asked for a "New Yorker-style cartoon," ChatGPT often adds the signature of a real cartoonist, Nieman Lab's Andrew Deck reports. He documented more than 15 New Yorker contributors whose pen names appeared on AI images, including Brendan Loper, whose signature sat on a viral cartoon about Dolly Parton that he never drew. "My name is my name," Loper said. "It felt very much like a violation of my personhood."

After Nieman Lab contacted OpenAI, ChatGPT began warning that some such requests may break its rules on similarity to third-party content, but it still signed some cartoons with real names. The New Yorker's owner, Condé Nast, has a licensing deal with OpenAI, but a spokesperson said it has never allowed an AI company to train on the cartoons.

Independent researcher Syed Anas Mohiuddin showed that a malicious instruction planted in one AI agent can travel to other agents inside a network, because agents connected through the Model Context Protocol (MCP) tend to trust each other. He calls it "protocol pivoting," and Google, Rapid7 and three other organizations have acknowledged flaws he found. The most serious, in a Google database toolbox for MCP, could make the tool send requests to internal systems on an attacker's behalf. Google has fixed it.

Some researchers say it is simply a form of prompt injection. Either way, the advice from Rapid7's Douglas McKee is blunt: treat anything an AI model passes to your tools "like input from a stranger on the internet."