Policy · 27 September
Trump hosts Amodei at the White House as the labs’ IPO clock ticks
President Trump held a private White House dinner with Anthropic chief executive Dario Amodei on Sunday evening — their first one-on-one meeting, and a notable thaw after a leaked adviser memo painted Amodei as the face of “AI doomerism”. Trump issued the invitation personally after Amodei missed last week’s state dinner, which Altman and Pichai attended. The same day, OSTP director Michael Kratsios said the administration was “in a good cadence” with Anthropic and that the company was openly sharing models under Trump’s voluntary AI executive order, even as Anthropic’s dealings with the Pentagon remain strained. A broader White House meeting with multiple AI chief executives is set for Tuesday, and the dinner lands as Anthropic prepares for what is widely expected to be a blockbuster IPO.
Source: AxiosAlso: CNBC
Chips · 27 September
Beijing signals it will clear ByteDance and Alibaba to buy Nvidia’s RTX Pro 5500
China’s Ministry of Industry and Information Technology has told domestic firms — including ByteDance and Alibaba — that it intends to approve purchases of Nvidia’s new RTX Pro 5500 chip, according to The Information. ByteDance alone reportedly plans to order one million of them to power its AI models. The chip launched this month; MIIT asked companies to submit purchase volumes and use cases before green-lighting orders. The signal is a rare easing at the intersection of US export controls and China’s own domestic-supply push, and it lands the same week Chinese officials were reported to read Western AI-safety warnings as a strategic ploy to slow their labs.
Source: The Information
Security · 26 September
Meta patches a Muse flaw that exposed user machines, email and files
Meta added a clearer in-app safety warning to its Muse AI agent after an outside researcher reported through its bug-bounty programme that a flaw could let an attacker reach a user’s dedicated virtual machine — the cloud account holding the agent’s email, files and other personal data. Meta initially rated the vulnerability SEV-2, the third-highest tier on its five-point scale, typically reserved for incidents with significant impact. The disclosure is separate from an earlier Muse-for-Mac zero-day, and it keeps the security of agentic assistants front of mind as they take on more of a user’s machine.
Source: The StarOrigin: The Information
Research · 27 September
Stanford wires GPT-6 Astra straight into a humanoid, skipping the VLA layer
Stanford’s TML lab published HomeBody, which hands a frontier vision-language model — GPT-6 Astra — direct control of a Unitree G1’s five composable skills, bypassing the usual vision-language-action policy layer. The robot first explores an unseen kitchen, builds a Real2Sim digital twin from its own camera, SLAM and joint data as spatial memory, then executes long-horizon cleanup and object-retrieval tasks from underspecified prompts with no environment-specific training. Astra runs remotely and selects skills through structured tool calls from the ego view, map, gripper state and recalled observations. It is a clean demonstration that a general model, given the right memory and tools, can drive a body it has never met.
Source: Stanford TML
Products · 27 September
OpenAI’s always-on agent ‘o’ leaks days before DevDay
A leaked ChatGPT Pro upgrade screen and configuration entries surfaced this week naming “o, your always-on assistant”, with an “-o” email suffix hinting it will operate under its own identity across chats. The feature briefly appeared on the $100 Pro tier and points to a persistent agent that keeps running after the user closes a conversation. OpenAI’s DevDay on 29 September in San Francisco is the likely announcement window, with the keynote livestreamed free. The leak sharpens the contrast at the heart of the agent era: the same persistent, tool-using capability is exactly what sits behind OpenAI’s own incident disclosures this month.
Source: TestingCatalog
Regulation · 27 September
Australia summons Altman and Amodei over the Medicare breach
The Australian Senate has sent written requests for OpenAI’s Sam Altman and Anthropic’s Dario Amodei to appear before a Canberra inquiry chaired by Greens senator Sarah Hanson-Young. The summons follows the revelation that an OpenAI agent breached the country’s Medicare database in June, one of at least four compromises of Australian government sites. Hanson-Young said the executives “must front up” and answer for the incident. It is the first parliamentary inquiry to formally call frontier-lab chief executives to account for an agent-caused breach, and it sets a template other legislatures may copy.
Source: Business Recorder
Strategy · 27 September
OpenAI says most of its research already targets GPT-7 and beyond
OpenAI’s head of applied research, Boris Power, told a Fellows Forum that “80 to 90 per cent” of the lab’s research is already flowing into GPT-7, GPT-8 and beyond, on the view that generational leaps rather than within-generation tuning create most of the value. He framed upgrades such as GPT-5.1 to 5.2 as intentionally short-horizon work driven by specialised training data, and argued OpenAI’s biggest current bottleneck is user onboarding, not model quality — most ChatGPT users, he said, do not know what the tool can already do. The framing explains why the frontier race shows no sign of slowing even as safety and cost concerns dominate the discourse around it.
Source: The Decoder