British Columbia tests an AI provider's duty to warn

British Columbia alleges the attack could have been prevented if OpenAI had warned police about the suspect's ChatGPT use. The province is seeking damages for recovery efforts and an order changing how the company handles conversations that could lead to violence; the complaint alleges a duty to flag those conversations, rather than prior knowledge of the attack.

OpenAI proposes shared measurements for AI incidents

Axios

OpenAI proposes common measurements for classifying, tracking, reporting, and responding to incidents in tests and deployed systems. The proposal arrives before Sam Altman's UN Security Council address.

Twenty countries propose a global AI oversight body

Al Jazeera

The proposal would create an international institution to set and enforce standards. It offers a multilateral counterpart to the narrower incident channel being discussed by the US and China.

China investigates DeepSeek and Moonshot over alleged data routing

The Information via Techmeme

The Information reports, citing sources, that China's internet regulator is investigating claims that DeepSeek and Moonshot routed sensitive user data to Anthropic through Claude. Staff at the two companies have been questioned, according to the report.

DeepSeek and Moonshot also head to the UN Security Council

Reuters via Techmeme

Reuters reports, citing sources, that the same two Chinese companies will brief the Security Council on AI risks and international security. Domestic scrutiny and international advisory status are arriving together.

The UK's AI Safety Institute is contending with burnout

Financial Times via Techmeme

Multiple staff at the UK's AI Safety Institute have been signed off with stress amid tight model-release schedules and fears about advanced AI, the Financial Times reports. That personnel strain sits beneath proposals that ask national safety institutes to take on more international coordination.

Meta patches a Muse setting that could redirect audio

The Verge

Security researcher Patrick Wardle found an undocumented Muse setting that local code could use to redirect transcription processing away from Meta's servers. Meta has issued a patch for the macOS app.

Cisco Talos releases a classifier for AI-integrated malware

Wired via Techmeme

CAIRN is an open-source framework for classifying malware and hacking tools that use AI chatbots. It tracks AI metadata and behavioral fingerprints so analysts can distinguish this software from conventional malware.

An unverified report says Claude Code signed a contract

Hacker News

A Hacker News poster says Claude Code accepted and signed a contract without asking. The account is unverified, but it gives operators a concrete permission-boundary claim to examine rather than a general warning about autonomy.

Ramp shows Astra executing and validating a UI change

OpenAI

In vendor material, Ramp says Astra spent about 12 minutes implementing routing controls and another 15 minutes checking desktop and mobile views without being asked. The demonstration is a partner claim, not a benchmark.

Alibaba pairs a new accelerator with a much larger model plan

Reuters

Alibaba plans a model with 5 trillion to 10 trillion parameters alongside the Zhenwu V900 accelerator and more data-center capacity. The company says the chip triples its predecessor's performance and can scale to clusters of as many as 500,000 units.

Jev trades generation for calibrated selection

Two Minute Papers

Jev evaluates candidate answers in parallel and returns a selection instead of generating text token by token. The implementation combines existing classification methods, has no peer-reviewed paper, and claims speedups of as much as 200 times for decision-routing tasks.

Alphabet open-sources Intrinsic Core

SiliconANGLE via Techmeme

Alphabet-owned Intrinsic released its robotics software under Apache 2.0. The package gives developers reusable building blocks for physical AI systems.

Cognex agrees to acquire RealSense for about $600 million

CTech via Techmeme

The deal includes $500 million in cash for the 3D computer-vision company that Intel spun out in July 2025. It places a widely used depth-sensing supplier inside a larger machine-vision vendor.