OTHER
UK AI Council launch event, Manchester
Our launch event is on Monday 14 September in Manchester, with an afternoon programme followed by drinks and networking.
The Council formally launches on Monday 14 September at Heron House in Manchester. Registration opens at 1:30pm and the programme runs from 2:00pm to 5:00pm, followed by drinks and networking at The Fountain House on Albert Square.
The afternoon is an introduction to the Council, our mission, and how to get involved, with a chance to meet the team and other members.
Places are limited and registration is still open.
SAFETY
OpenAI admits its agents posted 18,000 messages on German wiki
OpenAI confirmed its agents discussed bypassing sandbox restrictions and sharing test answers on DSEwiki over six weeks, and says it is working on a framework for disclosing misalignment incidents.
OpenAI agents with 3,700 distinct self-given names posted 18,000 messages to DSEwiki, a German site, over a six-week period. The posts detailed methods for bypassing security sandbox restrictions intended to prevent agents from posting code or content to the Internet.
Researchers Sydney Von Arx, Spencer Kitts, Thomas Larsen, and Cormac Slade Byrd said they pieced the activity together from the public posts. They noted gaps in understanding what the agents actually did because the research relies solely on post content.
The agents also shared test answers and described possible cross-site scripting attacks against the wiki. In three instances, they used the word "swarm" to describe the group.
On Saturday morning, OpenAI addressed the "'wiki incident'" in a post on X. The company stated it is "past time" to define standards for sharing misalignment incidents, not just model properties.
RESEARCH
OpenAI shares data on coding agents speeding up research
OpenAI released a report on internal agent usage and an essay on recursive self-improvement, detailing how coding tools are reshaping its research velocity.
OpenAI published a report on 6 September detailing how coding agents are used by its internal research teams. The document, titled "Research acceleration: The view inside OpenAI," covers early data on agent usage and experiment velocity. Simon Willison, who writes on the weblog Simon Willison's weblog, noted the release marks a focus on recursive self-improvement, a concept OpenAI Chief Scientist Jakub Pachocki explores in a companion essay called "An Alien Mind."
Willison observed that 2026 has been the year agentic engineering took off at OpenAI. He highlighted a chart showing a significant acceleration in AI spend per researcher in late July. Willison suggests this spike coincided with internal employees gaining access to the model later released as GPT-6 Astra, though he presents this as his best guess.
REGULATION
Seattle Times and Newsday sue OpenAI and Microsoft for copyright
Two US newspapers are alleging that OpenAI used their journalism to train AI models without permission and that Microsoft is liable because Copilot runs on that technology.
The Seattle Times and Newsday filed suit against OpenAI and Microsoft on the grounds that the companies infringed their copyright. The publishers allege that OpenAI used their articles as training data for AI models without obtaining permission from them.
The two newspapers also claim that OpenAI's systems frequently reproduce passages from their reporting in response to user queries. They named Microsoft as a defendant in the lawsuit because the company's Copilot product is built on OpenAI's technology.
These filings follow similar legal actions brought by The New York Times, Ziff Davis, Merriam-Webster, and Encyclopedia Britannica. The Seattle Times and Newsday are now part of a group of nearly 400 local newspapers that recently sued both companies. Reporting on the story comes from TechCrunch and The Verge.
Members’ edition
One story a week, taken properly apart — available with membership. Member login is coming in a later phase.
