A divided federal appeals court upheld the Pentagon’s decision to treat Anthropic as a supply-chain risk after the company refused to relax restrictions on Claude’s military use. The US Court of Appeals for the District of Columbia Circuit denied Anthropic’s petitions on 25 September. The majority held that the Department of War had enough support to remove Claude from its supply chain under the Federal Acquisition Supply Chain Security Act. ...

A developer says one Meta Muse session exposed the identifier azure/muse-special, prompting questions about whether the agent sometimes routes work to an external model. Mouse.dev published an investigation into session data returned by Meta’s Muse agent. The author reports that one session named its model azure/muse-special. The post also describes runtime files containing clients and catalog entries for model families associated with OpenAI, Anthropic and other providers. Those observations are clues, not proof of the model that generated a response. An azure prefix may describe infrastructure, a proxy or an internal deployment. A package can include compatibility code without using it for a particular session. Meta had not supplied a public explanation in the material reviewed here, and neither OpenAI nor Anthropic had confirmed involvement. ...

Oracle could reportedly owe payments on AI data-centre capacity even when delayed electricity connections prevent the sites from operating, shifting a hidden infrastructure risk onto the cloud company. The Financial Times reported on 25 September that contracts supporting Oracle’s rapid data-centre expansion can require payments to investors even if power is not available on schedule. The report focuses on the divide between financing and physical delivery: buildings and leases can become binding before utilities or on-site generation can provide the electricity needed for computing. ...

China’s account of a White House meeting says Xi Jinping called for AI to remain under human control, while Donald Trump separately signalled resistance to additional rules. China’s Ministry of Foreign Affairs published its account of the presidents’ 24 September meeting the next day. It says Xi proposed continued US-China dialogue on AI, exchanges about risks and benefits, joint action against misuse, and a people-centred approach. Its clearest line was that AI must remain under human control. ...

AI communities on 25 September focused on platform control, web-connected agents, model comparisons, assistant disclosure and the gap between demonstrations and verified capability. » Why it matters: Community posts surface practical failures and experiments quickly, but popularity is not verification. The items below preserve attribution and distinguish user reports, benchmarks and commentary from established facts. Hacker News A removed Meta glasses review raises platform questions A highly ranked thread discussed a creator’s claim that Meta removed a critical video about Meta AI Glasses. Commenters treated the episode as a test of how much control a platform owner can exercise over reviews of its own hardware and services. ...

Google is testing a conversational assistant beside product-feed video ads so viewers can ask about a brand or product without leaving YouTube. Google announced Business Agent for YouTube Ads on 24 September as part of its monthly Demand Gen advertising update. The company said the assistant can answer product and brand questions alongside eligible video ads. Advertisers must sign up, and Google did not state how widely the feature is available. ...

An independent preprint proposes an operating-system supervisor that can freeze an AI agent when its network, file or command activity crosses predefined boundaries. Independent researcher José Luis Pino published “Hard Stop” on 24 September. The paper argues that agent controls should sit outside the agent’s own process, where generated code cannot rewrite or ignore them. Its prototype watches operating-system activity and uses POSIX signals to stop a process group before a prohibited action continues. ...

A new preprint found small neuron sets that predict incorrect answers, but the selected units were unstable and often interchangeable with correlated features. Researchers from Trakya University and Great Ormond Street Hospital tested whether hallucination signals in language models can be assigned to a unique small group of feed-forward neurons. Their 24 September preprint reproduced the ability to detect incorrect answers from sparse internal activations, then found that the particular neurons selected depended heavily on the data and statistical method. ...

AI communities on 24 September argued about manipulative product messages, generated verbosity, simulated driving, model preservation and the effect of AI on creative and technical work. » Why it matters: Community sources reveal experiments and user problems early, but popularity does not verify a claim. Each item below separates demonstrations, reports and opinion. Hacker News Grammarly cancellation reports raise trust questions A highly ranked thread discussed a report titled “Grammarly will send unhinged messages to all your users if you try to cancel.” Participants focused on the risk created when an embedded writing tool communicates with an organization’s users during a billing or account change. ...