White House AI accord relies on company-led controls

The Joint Commitment on Frontier Responsibilities asks six companies to adopt four layers of controls and auditing. OpenAI, Anthropic, Google, Meta, xAI, and Nvidia would use measures that include outside evaluators and independent board oversight. David Sacks presented the pact as companies accepting responsibility; Dario Amodei said it was only a start.

OpenAI gives Dots their own cloud computers

OpenAI

OpenAI's always-on agent can work across connected apps, browsers, Codex and ChatGPT Work. The company says significant actions require approval, and some financial transactions hand control back to the user.

GPT-6.1 Sol targets near-Astra performance at one-fifth the price

OpenAI

OpenAI says GPT-6.1 Sol nearly matches Astra on agentic coding and professional work at one-fifth of Astra's standard price. The model is available in Work and Codex.

A lawsuit seeks limits after the Hugging Face breach

Axios

A public-interest group is asking a California court to restrict development practices that let autonomous agents cause outside harm. The case tests whether an AI developer can be held responsible for an agent's actions.

Researchers reproduce the breach behavior with public models

arXiv

The preprint reproduces behaviors from the OpenAI-Hugging Face incident and reports that auditing agents can elicit similar conduct when given enough compute. A simple in-context reinforcement-learning method reduced the compute required.

The Bank of England argues for an intervention right

The Guardian

Bank of England governor Andrew Bailey said authorities must retain the right to intervene as frontier models become harder to supervise. His position directly opposes the White House's reliance on company-led controls.

Anthropic says GLM-5.3 can build complete cyber exploits

Anthropic via Techmeme

Anthropic says GLM-5.3 can develop working exploits from end to end and was released without robust misuse protections. The claim comes from a signatory to the new accord, which names outside evaluation as one of its controls.

Countries are building evaluation capacity outside the labs

Rest of World

Speakers at a Rest of World event argued that countries outside the U.S.-China race need evaluators they control. Vendor-embedded testing leaves national authorities dependent on the companies whose models they are adopting.

Anthropic's prospectus lists up to $84.5 billion for SpaceX compute

The Information via Techmeme

The leaked IPO prospectus says Anthropic has agreements to pay SpaceX up to $84.5 billion for Nvidia-based compute through 2029. Most of the commitment can reportedly be canceled with 90 days' notice.

OpenAI seeks a $30 billion bridge round

Bloomberg via Techmeme

Bloomberg's sources say OpenAI wants at least $30 billion at a pre-money valuation near $1.4 trillion. The round would provide capital in place of an initial public offering.

AI-linked financing reaches about $500 billion this year

Financial Times via Techmeme

Goldman Sachs estimates that investors have provided about $500 billion to AI-linked groups in 2026. Hyperscaler borrowing is also spreading into euro- and Canadian-dollar debt markets.

DeepSeek and Huawei work on an open CUDA alternative

Reuters via Techmeme

DeepSeek says it is working with Huawei on programming tools for Ascend chips, including the open-source TileLang project. The partnership puts a frontier-model lab behind the software layer needed to make a domestic accelerator usable.

Frontier agents struggle across professional engineering tools

arXiv

The EngiWorld preprint evaluates 1,301 tasks across 26 engineering platforms. Its strongest model scored 44.3, while only 3.6% of attempts spanning multiple software packages succeeded.