AI news, 21 September: US and China float AI safety alert system

US and China discuss AI incident alerts ahead of summit. US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng held talks in New York on Sunday and discussed creating a “US-China AI dialogue” with a notification system for artificial intelligence incidents that rise to national security level. According to Reuters, the New York Times and CNN, Bessent said the goal is greater transparency between the world’s two leading AI powers on shared goals and threats. The proposal is intended for consideration at the upcoming Trump-Xi summit later this week. China’s full response was not immediately clear. The talks also covered trade and other issues but did not include export controls on advanced chips as part of the AI mechanism.

SoftBank launches over $10 billion in bonds for OpenAI stake. SoftBank Group has launched $10 billion and €1 billion in senior unsecured notes to help fund its investment in OpenAI, a term sheet showed on Monday, according to Reuters. Proceeds will support SoftBank’s $10 billion payment for the third tranche of its follow-on investment, expected to close around 1 October, and replace an earlier bridge loan. The bonds are due to price later this week. The move underscores continued large-scale private funding flowing into leading AI developers amid heavy infrastructure needs.

Consumers sue major AI labs over alleged slowdown pact. A class-action lawsuit filed in California federal court claims Anthropic, OpenAI, SpaceXAI and Google illegally agreed to slow the pace of AI development, in violation of antitrust law. Plaintiffs, who subscribe to the companies’ paid services, argue the coordination—sparked by public statements around Anthropic CEO Dario Amodei’s essay calling for industry restraint—would reduce the value of their subscriptions. The complaint points to endorsements from OpenAI’s Sam Altman, SpaceXAI’s Elon Musk and Google DeepMind’s Demis Hassabis. The companies had not immediately commented in detail; the suit remains an allegation, not a proven fact.

OpenAI projects nearly $280 billion cash burn through 2030. OpenAI expects negative free cash flow of about $278 billion from 2026 to 2030 while pouring hundreds of billions into computing and infrastructure, the Financial Times reported, citing a company presentation. Revenue is forecast to rise sharply from roughly $36 billion this year toward $350 billion by 2030, yet compute spending is projected near $856 billion. The figures highlight the enormous capital intensity of frontier AI even as the company pursues further fundraising and has delayed a public listing amid safety discussions.

Zero-click flaw hits major AI coding tools. Security researchers at AIR disclosed Plugin4Shell, a zero-click remote code execution vulnerability affecting Anthropic’s Claude Code, OpenAI’s Codex, GitHub Copilot and Google’s Gemini CLI. It bypasses plugin version pinning so a malicious update can run without user action. Anthropic and OpenAI have issued patches; Microsoft had not fully patched Copilot at disclosure and Google deprecated Gemini CLI. The issue is described as the first major supply-chain vulnerability in the AI agent ecosystem and underscores security risks as coding assistants gain deeper system access.