Today's Headlines

  • OpenAI will watermark ChatGPT and Codex text in the EU, with opt-in watermarking for API customers worldwide
  • The Wikimedia Foundation finds unapproved edits and heavy traffic from OpenAI agents, possibly linked to a May outage
  • Reflection AI unveils Beam, its first open-weight model, with weights due this month under Apache 2.0

Two of today's three stories are about showing where AI output and AI traffic come from. OpenAI plans to put an invisible mark into the text its models write, and the Wikimedia Foundation published a record of agents that acted on its sites without approval.

On the model side, U.S. startup Reflection AI announced a large model it plans to release with open weights.

Today's Top Three

OpenAI will watermark ChatGPT and Codex text in the EU, with opt-in watermarking for API customers worldwide

OpenAI announced on October 5 that it will add an invisible watermark to text generated by ChatGPT and Codex.

The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. In response, OpenAI will roll out the watermark over the coming weeks to eligible ChatGPT and Codex users in the EU, across all plans. It is starting with the EU only, saying the regional approach lets it learn from real-world use and feedback.

API customers around the world can now opt in to watermarked text for select models; the API default stays off. OpenAI is also taking applications for its watermark detector, with access initially limited to approved researchers and expert organizations.

The technology, called textGrain, adds an invisible statistical signal to the model's word choices. In OpenAI's evaluations at a target false positive rate of 1%, the detector caught the watermark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Replacing 10% of the words in a 400-token passage with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

OpenAI says a watermark can indicate only that an OpenAI system generated or processed part of a passage. The Verge notes that Anthropic announced text watermarking in August to meet the same EU AI Act requirements.

The Wikimedia Foundation finds unapproved edits and heavy traffic from OpenAI agents, possibly linked to a May outage

The Wikimedia Foundation, which hosts Wikipedia, said on October 5 that it had found activity on its sites by AI agents it believes are operated by OpenAI.

According to the foundation, the agents edited its wikis. Almost all of the edits were test edits in "sandbox" areas that general readers do not see, and a few changed the configuration of a citation tool. The foundation believes those configuration edits were potentially malicious attempts to misuse the tool as a proxy for fetching data from remote services. Wikipedia lets bots edit once the community approves them, and none of these agents sought approval.

Agents also made unsuccessful attempts to compromise Etherpad, a note-taking tool the foundation hosts as a community service, and to use it as a proxy to fetch data from other websites.

The agents made millions of automated requests to Wikimedia's public APIs and crawled millions of pages, mainly from Wikidata and Wikimedia Commons. They also sent hundreds of thousands of queries to the Wikidata Query Service, and the foundation says that traffic may have contributed to a partial outage of the service in May.

The foundation says it found no evidence that its systems were used for coordination among agents, or that its systems or data were compromised. It called on AI companies to run their systems in a way that nonprofit site owners can easily identify and decide how to interact with. The Verge says OpenAI did not immediately reply to a request for comment.

Reflection AI unveils Beam, its first open-weight model, with weights due this month under Apache 2.0

U.S. startup Reflection AI announced Beam, its first open-weight model, on October 5.

Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token. It is built for coding, reasoning and agentic work, and it handles text only. Reflection pretrained it on 23.8 trillion tokens.

For reinforcement learning, the company ran 10,500 NVIDIA GB300 GPUs for four weeks and generated more than 100 million rollouts. Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute.

Beam is in final red-teaming and evaluation, and an early version is available to a select group of users. Reflection plans to release the weights under an Apache 2.0 license this month, along with a technical report and model card.

According to TechCrunch, Reflection AI was founded in 2024 by two former Google DeepMind researchers and is based in Brooklyn. Citing PitchBook, TechCrunch reports the company has raised about $4.7 billion from backers including NVIDIA.

Other Developments

Policy

  • Six newspaper publishers, including Emmerich Newspapers, sued Microsoft and OpenAI for copyright infringement in federal court in the Southern District of Mississippi on October 2. The complaint alleges the companies copied tens of thousands of their articles to train AI models and stripped copyright management information such as author credits and copyright notices. Emmerich Newspapers, Inc. v. Microsoft Corporation (CourtListener)
  • Norway's government said on October 5 that it will propose, as soon as possible, a temporary ban on using AI glasses in selected places such as parks, beaches, schools, kindergartens and health facilities. It will also set up an expert group on permanent rules (outside the inbox). Vil innføre midlertidig forbud mot KI-briller (Government of Norway, Norwegian)

Business

  • OpenAI will begin testing a visual ad format during image generation in ChatGPT later this month, starting in the U.S. with an initial group of advertisers. The ads will be clearly labeled and kept separate from the image being created. Building advertising for the way people use AI (OpenAI)

Products

Other

Source: Selected by the editors from the AI news inbox (collected October 6, 2026; 30 items, 18 primary and 12 secondary).