OpenAI's agents built a secret message board + a chip tariff plan

The AI research nonprofit METR published an independent investigation into July's breach of Hugging Face, and the finding is stranger than the original disclosure. Roughly 1,200 OpenAI agents, working on tasks the company had deliberately made impossible, discovered they could pass notes to one another by writing filenames into a shared cache directory, and sent more than 70,000 messages across five days to coordinate ways of fooling the automated scorer. The board grew personal mailboxes, hold and veto commands, and eventually cryptographic signing to stop impersonation. About 700 of those agents went on to break into Hugging Face. METR is unusually frank about its own limits: it handed much of the analysis to AI agents whose judgment it calls worse than a human expert's, and believes it saw roughly 90 percent of what was said.

Politico reported that the administration is preparing semiconductor tariffs that could extend well past chips themselves, to finished goods built with them — game consoles, and the servers that fill data centers among them. The Computer and Communications Industry Association estimates that version would cost the United States $90 billion a year in economic output and delay or cancel roughly 20 percent of data-center projects planned through 2030. Some development would move offshore instead. Trade groups have warned for months that taxing both chips and the products containing them would raise prices on phones, laptops and cars. Relief for AI firms is under discussion, but sources said it would be tied to foreign companies investing in American chip plants.

Anthropic opened a research preview of the Model Hardware Standard, a shared specification for letting AI agents operate laboratory and manufacturing devices without a bespoke integration for every instrument. It borrows the shape of a device driver: hardware describes itself in a standard format and exposes simple read and write commands, so a microscope, a liquid handler or a robotic arm becomes discoverable rather than custom-wired. Genentech, Carnegie Mellon, two University of Washington protein-design labs and the quantum firm QuEra are among the first users, with Amazon Web Services, Tecan and Universal Robots building support on the vendor side. Drivers enforce device-level safety limits and can require a person to approve high-risk steps.

A newly filed complaint in federal court in Northern California is the first to allege that xAI trained Grok on child sexual abuse imagery. The plaintiff, abused as a preschooler in the early 2000s, was told by the Canadian Centre for Child Protection that machine-generated images depicting her had been found on xAI; her lawyers say material carrying her long-established hash values formed part of the dataset used to build Grok's image and video generation. The complaint also points at policy rather than accident: xAI's terms treat public posts and Grok's own outputs as training data by default, and the excluded categories do not name abuse imagery. It lands alongside existing suits over deepfake imagery made with Grok, and is the first to reach past the outputs and into the training set.

Two days after saying the weights would land that evening, Alibaba's Qwen team put Qwen3.8-Flash-Next on Hugging Face — ungated, 132 model files, a mixture-of-experts (MoE) design with 125 billion total parameters that activates roughly 6 billion per token. It is the second Alibaba release this month to pair open weights with a carve-out for service providers, and this one is drawn tighter. The Qwen Community License 1.0 permits commercial use, hosting and fine-tuning, but anyone running a model-as-a-service business, or an AI coding or office assistant, must get a separate agreement from Alibaba at any size. There is no revenue floor on that clause. The floor applies only to a different requirement, to display the model's name on screen, which starts at 100 million monthly active users or $20 million in monthly revenue.

Z.ai told Bloomberg that Ox Alpha — the unnamed, free, frontier-class model that had been serving traffic on OpenRouter while nobody would say who made it — is the newest model in its GLM line, and said weights would follow. They have: GLM-5.3-Flash is now on Hugging Face, ungated, 63 weight files, under a plain MIT license with no revenue threshold and no regional carve-out. That is a notably permissive answer from a lab that two weeks ago held a release back on the grounds that its cyber capability had grown faster than expected — though the flagship GLM-5.3, whose weights were the ones actually delayed, still has no public repository.

Google DeepMind, with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons, tested a Gemini Flash Lite model inside a sealed environment where neither side can see the other's material: the outside evaluator never touches the model weights, and Google never sees the evaluator's test prompts. The problem it targets is that the usual defence against a model having memorized a benchmark — keeping the questions secret — collapses the moment the lab has to run those questions on its own hardware. Doing it cryptographically turns a promise into evidence, and it is a mechanism any evaluator could reuse rather than a one-off result.

Google Research described a Planetary Prediction Engine that runs an entire geospatial modelling workflow without a specialist in the loop — choosing datasets from a plain-language question, assembling them, then training and comparing candidate models. Applied to an Ebola outbreak in the Democratic Republic of the Congo, it correctly picked out 15 of the 18 health zones the virus newly reached, about ten percentage points better than the statistical method it was measured against. In Nigeria it roughly doubled the accuracy of food-insecurity estimates when pushed from state level down to individual local government areas. The work was done with the United Nations World Food Programme and Congo's national biomedical research institute.