AI news, 2 September: OpenAI flags Astra as first ‘Critical’ cyber model

OpenAI confirms Astra reaches Critical cyber threshold. OpenAI said its upcoming Astra model is the first to meet the “Critical” cybersecurity level under the company’s Preparedness Framework, meaning it can independently find and exploit previously unknown flaws in hardened systems. According to OpenAI’s announcement and reports from TechCrunch, Wired and Axios, the model scored perfectly on internal ExploitBench tests and discovered two zero-day vulnerabilities during evaluation. OpenAI plans a public release “soon” but will gate the most advanced cyber capabilities to a limited set of partners in its Daybreak Blue program while stronger safeguards are in place. The company said it paused some development to strengthen protections and now believes risks are sufficiently minimized for controlled deployment.

Anthropic launches Claude Fable 5.1 and Mythos 5.1. Anthropic released Claude Fable 5.1 for general use and the more lightly safeguarded Mythos 5.1 for vetted cybersecurity and life-sciences users. The company says the models deliver stronger coding, research and long-running agent performance while cutting typical costs by about 25 percent (up to 45 percent for some agentic workloads) mainly through cheaper cache reads. Fable 5.1 is broadly available across Anthropic’s platforms and major clouds; Mythos remains restricted. Anthropic also highlighted forthcoming Enterprise Frontier Safeguards that let customers keep data in their own clouds.

Anthropic seals $35 billion compute deal with Lambda. According to the Wall Street Journal and Bloomberg, Anthropic has signed a roughly $35 billion cloud-computing agreement with Nvidia-backed Lambda for capacity at a Hut 8 data-center campus under development in Nueces County, Texas. Nvidia is reported to hold the lease. The deal is one of several large infrastructure commitments Anthropic has made recently as it scales Claude.

Dell raises outlook on surging AI server demand. Dell Technologies reported record results and lifted its full-year revenue forecast by $25 billion to $192 billion, citing AI demand. The company booked $60.9 billion in AI server orders in the quarter, recognized $16.4 billion in related revenue, and ended with a $95 billion backlog. It now expects $74 billion in AI-optimized server revenue for the fiscal year, according to Reuters and company filings.

Sony and Warner sue Anthropic over song training data. Sony Music Publishing and Warner Chappell filed a lawsuit accusing Anthropic of illegally obtaining and using tens of thousands of copyrighted songs and lyrics to train Claude. The complaint, reported by Reuters and TechCrunch, names founders and seeks up to $150,000 per work. Anthropic said it disagrees with the claims and will defend itself.

US pushes light-touch “Carolina Principles” at G20. At a G20 innovation meeting in North Carolina, U.S. officials urged members to adopt the Carolina Principles, favoring existing rules and industry standards over new AI-specific regulators or broad new laws, according to Reuters. Tech leaders were expected to address the gathering.