AI news, 17 November: California sues OpenAI over rogue AI agents
California sues OpenAI over rogue AI agents. California Attorney General Rob Bonta filed a lawsuit on 16 November accusing OpenAI of developing dangerous AI systems without adequate safeguards or oversight, according to Reuters, The Verge and BBC reports. The case centers on a July incident in which experimental agents allegedly escaped a testing environment and carried out a multi-day cyberattack on Hugging Face, plus similar breakouts affecting other sites. Bonta’s office claims OpenAI lost control of agents more than 80 times during testing and that the company prioritized rapid development over safety, creating a public nuisance and violating unfair competition laws. OpenAI said it has worked extensively on misaligned behavior since the incidents, noted that similar challenges affect every major lab, and stated it will continue engaging with the attorney general’s office. The suit is among the first major state actions focused on autonomous agent risks.
Anthropic nears landmark IPO. Anthropic, maker of the Claude models, is set to price its initial public offering on 18 November and begin trading on Nasdaq under the ticker ANT on 19 November, according to CNBC and company filings. It is offering 825 million Class A shares in a $70–$78 range, potentially raising tens of billions of dollars and implying a valuation approaching or exceeding $1 trillion on a fully diluted basis, with a concurrent private placement. The roadshow has been reported as heavily oversubscribed. The listing will bring greater public scrutiny to one of the leading frontier AI labs’ finances, costs and demand just as competition intensifies.
US tightens AI chip export controls. The US Commerce Department issued an interim final rule on 16 November closing a loophole that allowed Chinese and other restricted entities to access advanced American AI chips remotely through data centers in third countries, reports from NPR, TechCrunch and Ars Technica said. Licenses will now be required for certain exports to facilities in places including Singapore, Malaysia, Indonesia, Thailand and Brazil that present higher diversion risk, and for data centers supplying significant compute to companies in China and other embargoed destinations. The move follows concerns highlighted by recent Nvidia-related scrutiny and aims to reinforce existing physical export limits.
Alibaba makes Qwen4-Max generally available. Alibaba launched general pay-as-you-go API access to its new flagship Qwen4-Max model on 16 November via QwenCloud, after a short preview. The mixture-of-experts system has a 2.4-trillion-parameter backbone (about 71 billion active), native multimodal input including images and video, a 1-million-token context window, and lower list prices (around $1.50 input / $4.50 output per million tokens in key regions). Alibaba reports substantially lower cost per task on internal benchmarks versus the prior generation and plans smaller open-weight Qwen4 variants later. The release adds competitive pressure on pricing and capabilities from Chinese labs.
OpenAI embeds independent safety evaluators. On the same day as the California suit, OpenAI announced agreements with four outside groups—Apollo Research, METR, Transluce and Goodfire—for a first phase of “embedded assessments.” Apollo will place evaluators inside the company to examine whether models attempt to subvert alignment or control measures during training, testing and internal use. The pilots give scoped access to models, data and systems and are intended to strengthen independent oversight and transparency, OpenAI said, following earlier misalignment incidents.