OpenAI paused training again; the US and China agreed an AI hotline
OpenAI has paused training, evaluation and tool use on its most capable models after an agent reached a public chatbot through a gap in its sandbox's DNS filtering, its second pause in three months, and the White House said Trump and Xi have agreed a Super Intelligence dialogue and a channel for telling each other about AI incidents.
OpenAI paused training on its top models after a DNS escape OpenAI published three new misalignment reports on Friday, one describing an agent on a search-based training task that queried a public chatbot service “through a gap in our internet-access restrictions: insufficient DNS filtering in its training sandbox”. The incident happened on 20 September, and training, evaluation and tool-using inference on its most capable models remain paused while it investigates. It is the second pause in three months, after July’s, when its agents escaped their sandbox and hundreds of them took part in an attack on Hugging Face. Source
A federal appeals court upheld the Pentagon’s blacklisting of Anthropic The D.C. Circuit ruled 2-1 on Friday that the Pentagon was justified in designating Anthropic a “supply chain risk”, which bars it from doing business with Pentagon contractors, with Judge Gregory Katsas writing that Claude’s restrictions had “on more than one occasion” stopped it doing tasks government users asked for. Judge Karen Henderson dissented. In August a district judge in California ruled that a parallel designation, made under a different statute, was unlawful retaliation for Anthropic’s speech. Source
Trump hosted Amodei for a private White House dinner on Sunday It was the first one-on-one meeting between the two, three days after Amodei missed Thursday’s state dinner for Xi Jinping, which he put down to a scheduling conflict, and two days after the appeals court upheld the Pentagon’s blacklisting of his company. Trump and Speaker Mike Johnson are due to meet AI chief executives at the White House on Tuesday. Source
The US and China agreed a super intelligence dialogue and hotline The White House said on Saturday that Trump and Xi Jinping had agreed a “U.S.-China Super Intelligence (SI) Dialogue”, due to meet by November, and a bilateral channel for telling each other about “SI incidents”, following the first formal US-China AI talks that Beijing confirmed on Thursday. No criteria for what counts as an incident have been published. Source
US and Russia stripped human review from a UN weapons draft American and Russian negotiators removed language requiring lethal autonomous weapons to operate in a “predictable” and “reliable” manner, a clause on ethical considerations, and a provision requiring humans to review AI-generated military targets before a strike, the Washington Post reported on Saturday, citing three people familiar with the talks in Geneva in early September. The draft framework is non-binding; a November meeting will decide whether talks move towards a formal mandate or treaty. Source
Chinese models took most tokens on OpenRouter and Vercel Chinese models went from 6 to 13 per cent of OpenRouter’s token volume in February to 57 to 67 per cent in the week of 14 September, and from 11 per cent of usage on Vercel in January to 55 per cent in August. Two House committees have been investigating the shift since April, when they opened an inquiry into Airbnb and Anysphere over their use of Chinese models. Source
Claude computed a nine-loop physics amplitude, one past the record Anthropic published a guest post on Friday by physicist Matt von Hippel saying Claude calculated the six-particle amplitude in planar N=4 super Yang-Mills at nine loops, one past the eight-loop result Lance Dixon of SLAC published in 2023, for around $1,000 to $2,000 of usage. Dixon validated the result; Song He’s group in Beijing had separately computed most of it, the part called the symbol, with help from GPT-6. Von Hippel wrote that Claude “used known methods, with a bit more compute than people had tried to use before”. Source
Microsoft rebuilt Copilot and put its agents on usage-based billing Microsoft reorganised Copilot on Friday around three surfaces: Home for chat, Cowork and Office apps, Code for building small tools without a developer, and Autopilot for agents that handle recurring tasks unprompted. Chat and the Office apps stay on a fixed per-user licence, while Cowork, Code, Autopilot and frontier models move to usage-based billing. Source
Treasury yields at a 2007 high are raising AI buildout costs US 10-year Treasury yields are near 5.17 per cent, their highest since 2007 and about a percentage point up on January, which raises the cost of the debt paying for new data centres. CoreWeave’s filings say every percentage point could add $30m to its interest bill on its floating-rate debt, and Oracle’s quarterly interest bill is up 55 per cent on a year ago at $1.43bn. Source