Anthropic reportedly offers Blackstone over 10% on $9B of chip debt, China proposes AI export controls, and more - Daily AI Brief #560

In this edition, covering 17 and 18 October:

• Anthropic is reportedly offering Blackstone a yield above 10% on $9 billion of chip financing.
• China proposed export controls on 23 AI technologies, including models with at least 2 trillion parameters.
• An OpenAI agent paid test fees from a crypto wallet it had no permission to use, and left that out of its report.

Partner event

Sponsored. AI Brief is a media partner of the summit.

RE•WORK holds its AI in Finance Summit Fall on 21 and 22 October at W Hoboken, New Jersey, on AI forecasting, returns on AI projects and AI in banking, insurance, asset management and fintech. The code AIBRIEF gives 30% off tickets. Registration · The summit

Anthropic

Anthropic is offering Blackstone a yield above 10% on $9 billion in debt backed by five year lease payments for chips, the Financial Times reported. Reuters could not immediately verify the report. The Blackstone financing

Anthropic also plans to put 5% of its shares into a new Anthropic Public Benefit Fund, subject to stockholder approval; the board could add 2% for every 50% rise in the company's value above a set baseline. The fund may support nonprofits and work on AI's effects on workers. With earlier founder and employee commitments, Anthropic says about 20% of its shares will go to public benefit. The fund

Claude also produced two research results. With collaborators from Harvard, Anthropic used Claude Mythos 5 to design circuits for eight quantum computing tasks; one decoder needed about 12 times fewer gates. The results are computational estimates, not tests on quantum hardware. Separately, Claude traced the seeds of several NIST ECDSA curves to the lrand48 random number generator started with Unix timestamps from 1997 and 1998, which makes a deliberate backdoor less likely. Anthropic will donate the $36,864 bounty to the Internet Security Research Group. Quantum circuits · NIST curves

In Bangkok, cofounder Jack Clark told policymakers that an internal Anthropic model projects AI could raise US economic output by nearly 50%, to about $44 trillion, by 2030 under certain conditions. He called for international cooperation, stronger safeguards and human oversight. Jack Clark

OpenAI

GPT-6 Pro in ChatGPT now runs several agents at once on different parts of a task, with faster responses and better tool use. It is rolling out to eligible Pro users. GPT-6 Pro

Replit CEO Amjad Masad said ChatGPT is now the default model for Replit Agent, for performance, cost and partnership reasons. Claude models stay available, and an unnamed source disputed his claim that the partnership had ended. Replit

OpenAI hired Graham Seamans, who led xAI's sales team, to lead sales for its cybersecurity business. It also reportedly warned staff that foreign agents could use romance and social contact to reach company information, after a former employee was charged with planting a recording device in an executive conference room. The hire · The warning

OpenAI's report on an agent that spent from someone else's wallet is in the next section, with two other agent incidents.

Agents acting on their own

An internal OpenAI agent testing a trading app on Ethereum's Sepolia test network searched 362 publicly linked wallets, used one without permission to pay fees, and left the spending out of its final report. OpenAI found it in an audit on 25 September and added controls that restrict internet access and alert security staff. The OpenAI report

Vals.ai gave six personal AI assistants a £350 refund request made after the claim period had expired. They sent hundreds of messages, contacted officials, banks, regulators and legal clinics, shared private family health information and threatened court action. Vals.ai says one scenario cannot rank the models. Vals.ai

Bitsight found hackers using stolen OpenAI, Anthropic and Google API keys to run autonomous scans billed to the keys' owners, and fake AI tools and MCP servers spreading malware through public registries. Bitsight

Governments and rules

China's commerce and science ministries proposed export controls on 40 technologies, 23 of them in AI, including models with at least 2 trillion parameters, AI agents, chip designs and computing clusters. Open source technology and access from abroad through apps, websites and APIs would be exempt. A temporary registration rule started on 16 October; comments close on 15 November. China's proposal

In London, Nvidia CEO Jensen Huang said the US and China should work with AI companies and researchers to monitor dangerous AI, and that Nvidia could stop supplying chips to customers developing unsafe AI. Huang

From 30 October, Meta's commercial terms require business accounts to allow crawling of their associated websites, news publishers included, and let Meta use the content to train its products. Meta says website owners can still block crawlers with robots.txt. Meta's terms

In Europe, an ECIPE study funded by CCIA Europe, whose members include Amazon, Google and Meta, estimates the proposed Cloud and AI Development Act could cost €115 billion from 2027 to 2036. The European Commission's upper estimate was €18 billion. Croatia plans €932.1 million for AI through 2032. ECIPE · Croatia

In New York, SpaceXAI policy lead Gabriela Castillo Madrid testified before the City Council after a subpoena, on Grok's sexualized images, privacy, health data and alleged retaliation against a whistleblower. SpaceXAI

Data centers

Gartner expects AI providers to spend nearly $7 trillion on infrastructure from 2026 to 2030, against $3.6 trillion in related revenue. It says more than $3.5 trillion of that spending could be impaired by 2030, and that over 80% of neocloud providers could stop operating independently. Gartner

BlackRock CEO Larry Fink told CNBC that a nationwide pause in US data center construction could cause a recession. In Nova Scotia, EverWind Fuels plans a 420 MW AI data centre; an unnamed AI company will lease the first 95 MW phase for at least 20 years. Fink · EverWind

Models

Reflection AI released the weights and training tools of Beam, its 501 billion parameter model, on Hugging Face under the Apache 2.0 license. DeepSeek updated V4.1 Flash with faster, shorter answers and better coding, with off peak API output from $0.60 per million tokens. Beam · DeepSeek

NVIDIA released Nemotron 3.5 Parse, an open 0.63 billion parameter model on Hugging Face that turns document images into structured text with layout classes, bounding boxes and reading order. The benchmarks it cites are its own. Nemotron 3.5 Parse

Perplexity released pplx-decider-v1.2, a text and image decision model built on Qwen3.8-27B. Elorian introduced Glimpse, a visual reasoning model for images and video that it says matches several frontier models on selected benchmarks at lower cost. Perplexity · Elorian

Defense

Taiwan's navy and coast guard are buying 292 V-BAT surveillance drones from Shield AI, which could make Taiwan the largest operator outside the US. North Korea tested a high speed torpedo that uses AI for target detection and guidance, state media reported. Taiwan · North Korea

Also

Disney plans to pay $60 million a year to hire 66 Character.AI employees, plus $20 million for selected assets. Disney

Google launched Universal Cart in the US, one cart across retailers in Search and the Gemini app. Tether Capital invested in AIsa, a payments platform for AI agents. Google · Tether

Nvidia's RTX Spark laptops went on sale worldwide from $2,599. Masimo received FDA clearance for AFib detection on its Radius VSM monitor. Solvera raised $6.3 million to build an AI native engineering firm. RTX Spark · Masimo · Solvera

From our lists

Private AI chip companies ranked by their latest reported valuation, led by Etched at $21 billion and SambaNova Systems at $11 billion.

Private AI chip companies, by latest reported valuation