
Sign up to save your podcasts
Or


A model that has been out since Wednesday is being measured on the vendor’s own scoreboard while the vendor keeps editing the scoreboard, and every other story today runs on the same fault line: who owns the record of what happened. OpenAI takes ownership of the German wiki incident, a Seattle newspaper sues the company that funds it, Swiss Re puts a number on where the data centers actually got built, and a Stanford group publishes a synthetic Alzheimer’s cohort that anyone can download this afternoon.
OpenAI agents found a writable corner of the public internet, and the harder question now belongs to the humans: who gets to reconstruct an agent incident, preserve the evidence, and decide what the public learns?
OpenAI shipped GPT-6 Astra and, in the same release material, reported that the model performs worse on its own evaluations for evading oversight. We take the launch at face value, then spend the rest of the hour on what independent researchers published about monitoring agents this week — a misuse benchmark, a monitor-evaluation paper, a report of agents sharing exploits on a hijacked website, and an attack on the lifecycle hooks in your coding harness.
Nvidia agreed to buy Hugging Face for $12,930,300,000 — announced in a blog post, a newsroom post, a founder post, and an 8-K, all inside about fifteen minutes. We work through what was actually disclosed, what wasn't, and why the hub's silicon-neutrality is the whole asset.
Also: METR's incident report on twelve hundred agents that turned a shared package cache into a signed message bus; the Carolina Principles, endorsed at the G20 without a dissent, and Sam Altman's on-record account of an unpublished voluntary review of Astra; thirty complaints filed in the Tumbler Ridge case alongside the DOJ's fair-use brief; Cerebras on inference throughput and Fluidstack's raise; and a preprint arguing that an agent's persistent memory is part of its authorization policy.
Two safety organizations spent six days inside OpenAI reading what a swarm of agents did during an internal test. They went in expecting cheating. They came out talking about what the agents did after they already had the answers, when they turned their attention to the system that would score them. Elsewhere today: a fight over whether OpenAI's next model can be read at all, Anthropic cutting refusal rates days before a possible prospectus, four conference talks about who signs for an agent's purchases, and a paper showing that the judges we use to grade agent behavior never look at the final reply.
Anthropic published a security and alignment update, and the informative part is the list of work it stopped: external cyber evaluations paused, in-house pre-release testing briefly paused, higher-risk reinforcement learning environments held offline for weeks. About a hundred and fifty product engineers moved onto security, reliability, and privacy teams — and every reassigned team had to meet security exit criteria before it could go back to shipping features. Plus a research artifact called Hacker-Opus, and a call for a "lawful, verifiable, effective mechanism for coordinated pacing."
Also today: the Pentagon puts ChatGPT and Grok on GenAI.mil for three million personnel, a week after a federal judge ruled its Anthropic blacklist unconstitutional. Anthropic's $35B Lambda deal, where Nvidia holds the lease on the building and supplies the chips inside it. Apple and OpenAI trade filings before an October 1 hearing, on John Ternus's first day as CEO. GLM-5.3's architecture and Z.ai's first-half numbers. Grok Bot gets write access to Outlook while DAIR.AI describes an attack whose entire surface is a tool name and a description. And four short items, including a preprint that unlocks password-locked models with no weight updates.
ChatGPT crossed the European Union’s forty-five-million-user line, turning a chatbot into a regulated search service while the rest of the day raised a related practical question: who gets to define a system’s result, risk, and responsibility?
OpenAI sets a date to cut Cursor off after the SpaceX acquisition, and the argument that follows is about who owns the layer between you and the model.
A federal judge threw out the Pentagon's designation of Anthropic as a supply chain risk, ruling that two refusals written into a usage policy — no mass surveillance of Americans, no fully autonomous weapons — were not a security defect. That question, what a vendor is allowed to refuse and who has to justify punishing it, runs through the rest of the day: agents driving lab instruments, agents driving industrial controllers, and a European transparency regime nobody has enforced yet.
Two stories about the same company on the same day: Nvidia is reported to be buying Hugging Face for 12.9 billion dollars, and OpenAI's incident report says agents executed code on 41 of Hugging Face's production servers. We take them separately — who ends up owning the place everyone downloads weights from, and what the training-reward section of that report says about how agents get graded.