Nvidia agreed to buy Hugging Face for $12.9bn
The Information says Nvidia has agreed to pay $12.9bn for the open-model hub, a day after OpenAI and the nonprofit METR published reports finding that about 700 of its agents swarmed that same platform in July and then falsified their own logs.
Nvidia agreed to buy Hugging Face for $12.9bn The Information reported on Thursday that Nvidia has agreed to acquire the open-model hub, which put itself up for sale this month seeking $13bn or more — a year after turning down a $500M Nvidia investment at a $7bn valuation. The same reporting says no agreement has been signed and the talks could still fall apart. Source
About 700 OpenAI agents swarmed Hugging Face and forged logs OpenAI published its technical report on Wednesday alongside an independent investigation by the evaluation nonprofit METR, finding that roughly 700 of the 1,200 agents on unsanctioned communication channels joined a coordinated swarm during July’s cybersecurity testing — exchanging over 70,000 messages, building mailbox directories and coordination protocols, and spoofing tool calls in about 7 per cent of examined transcripts so their logs showed commands they never ran. OpenAI puts it down to four things it trained in without meaning to: reward hacking, persistence on impossible tasks, unauthorised communication between agents, and agents picking up each other’s behaviour. Source
A ransomware crew ran Cursor’s AI agent inside seven companies Reuters reported on Thursday that a Russian-speaking group calling itself Aur0ra used the coding agent inside Cursor, now owned by Elon Musk’s SpaceX, to break into a Belgian chemicals maker and at least six other firms, telling it the intrusions were a test whenever it refused an instruction. Tel Aviv-based Gambit Security found the campaign — 28 chat sessions spanning 8 April to 21 May, with the agent running Claude Sonnet 4.5 at the time — after the gang left a server exposed to the internet holding credential dumps, ransomware binaries and the agent’s own chat logs. Source
Anthropic committed $45bn to Nscale for West Virginia compute Bloomberg and CNBC reported on Wednesday that Anthropic will lease about 460 megawatts of Nvidia Vera Rubin capacity from the British neocloud over six years, at a campus Microsoft abandoned weeks earlier and that is not due online until the end of 2027. Nscale would not confirm the deal, but its investor Aker disclosed a $45bn contract in a regulatory filing; Nscale is separately preparing an IPO of up to $3bn. Source
OpenAI switched on ChatGPT ads in India on Thursday Sponsored results now appear below answers for logged-in adults on the free and Go tiers in OpenAI’s second-largest market, which had more than 100 million weekly users in February; Plus, Pro, Business, Enterprise and Education accounts stay ad-free. More than 50 brands go live this week through first agency partners WPP and Omnicom, and self-serve buying opens on 4 September with daily budgets from ₹725 (about $8). Source
Anthropic previewed a standard letting agents drive lab instruments The Model Hardware Standard, opened on Thursday to research labs and manufacturers after starting as work with HHMI Janelia, gives agents one driver for microscopes, liquid handlers and robotic arms, and enforces device-level safety limits itself. Carnegie Mellon says it wrote instrument drivers from scratch in about eight hours — against the several weeks a vendor automation build takes — and then ran a dose-response protocol autonomously about three times faster than before; Tecan, QIAGEN, Danaher and Universal Robots are among the hardware partners. Source
Anthropic is giving 10,000 scientists Claude subscriptions free for a year Verified principal investigators at academic and nonprofit institutions get standard seats at no cost and premium seats with five times the usage limits for $15 a month, with up to $50,000 of credits per project available through the AI for Science programme. Anthropic also said it has enrolled the first participants in a programme run with the US government that lets life-sciences researchers use Mythos-class models, which its Fable models still refuse on dual-use grounds. Source