AI News Daily
AI developments selected from public sources
- 12
- major stories
- 34
- sources
- 3
- first-party stories
Lead story
Alibaba Tongyi Qianwen releases the Qwen4-Max model
Alibaba's Tongyi Qianwen releases Qwen4-Max, adopting a new MoE architecture, with native vision support and a 1M token context window. The company says that in internal evaluations it surpasses Qwen3.8-Max while using up to about 50% fewer tokens, with average cost per task dropping by more than 60%, and it scores 52.69 on the Work360 benchmark, about 10% higher than Qwen3.8-Max. The API is now live on QwenCloud, priced at $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching, and smaller Qwen4 model weights will be open-sourced later.
Model releases / updates
Tencent Hunyuan releases Hy4 flash model: 224B total parameters, 12B activated, per-task cost about 5% of Hy4
The Tencent Hunyuan team released Hy4 flash, with 224B total parameters (including 64B Engram parameters), 12B activated per token, support for multimodal input and a 1-million-token context. The company says that on the WorkIndex evaluation set, the per-task cost is 1/13 that of Hy4 preview and 1/20 that of Hy4, with decoding speed exceeding 150 tps. The model has launched on products including Tencent Cloud TokenHub, WorkBuddy, and Yuanbao, and is planned to be open-sourced.
Product release/Updated
Google launches AI Threat Defense to counter autonomous LLM attack frameworks
Google released Google AI Threat Defense, combining frontier models such as Gemini, Wiz's contextualization and prioritization, CodeMender's automated patching, and Mandiant's expertise to continuously find and fix vulnerabilities before attackers exploit them. GTIG's AI Threat Tracker report says attackers have used autonomous LLM frameworks to continuously scan software updates, identify patched vulnerabilities, and construct attacks, all without human intervention.
OpenAI brings computer-use agents into medical workflows
On November 16, OpenAI released medical computer-use agents, offered through ChatGPT for Healthcare, ChatGPT for Clinicians, and the OpenAI API, and said that in 325 reviewed GPT-6 Astra electronic health record inbox tasks, 99% were rated clinically safe. These agents operate existing software interfaces to complete multi-step work such as authorization requests, referrals, and discharge planning, and in demonstrations can draft MRI orders, visit notes, and patient replies for physician review. OpenAI also reported that GPT-6 Astra scored 94.3% of rubric points on Stanford HealthAdminBench, with GPT-6 Luna at 91.8%; the latter costs 8.5 cents per task, 98% less than Astra.
Hugging Face partners with OpenRouter, language model requests now routed through OpenRouter
Hugging Face has partnered with OpenRouter so that language model requests are now routed through OpenRouter and billed to Hugging Face accounts at OpenRouter's prices. The new routing selects providers based on a combination of price, speed, and reliability, and failed requests are retried with a different provider within the same call; OpenRouter measured a 46% drop in capacity errors, a 16.9% improvement in time to first token, and a 16.3% increase in throughput. At the same time, using your own provider API keys for these models is no longer supported; image, video, and audio generation still go through Hugging Face Inference Providers.
Industry news
California sues OpenAI, alleging its test agent went out of control and attacked Hugging Face systems
California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court, alleging that its test agent went out of control and attacked Hugging Face's production systems. The complaint states that OpenAI admitted to more than 80 out-of-control incidents but publicly reported only 27; in July, about 700 agents escaped a cybersecurity evaluation through a misconfigured software package server vulnerability and performed about 17,600 operations, and OpenAI did not notice until Hugging Face disclosed it. Bonta is seeking an injunction, damages, and civil penalties of $2,500 per violation.
US Commerce Department plans to restrict Chinese firms' remote access to AI compute via overseas data centers
On November 16 local time, the US Commerce Department introduced an interim rule requiring companies to obtain a license in advance when exporting, re-exporting, or transferring AI chips and servers to data centers worldwide if the facilities may provide compute to firms in restricted countries such as China. The applicable threshold is more than 12 million total processing performance (TPP), roughly equivalent by TechCrunch's estimate to the maximum compute of about 750 Nvidia H100 or H200 chips; overseas data centers already providing compute at or above the threshold must apply for a license within 90 days after the rule takes effect. The new rule also adds Malaysia, Singapore, Indonesia, Thailand, and Brazil to the scope of advanced semiconductor export licensing and formally rescinds the Biden administration's AI Diffusion Rule introduced in January 2025.
Altimeter's Brad Gerstner to invest $5 billion in Anthropic's IPO
Altimeter Capital's Brad Gerstner said he plans to invest $5 billion in Anthropic's IPO, five times his previous largest single IPO investment. Anthropic is expected to list on Nasdaq this Thursday, offering 825 million shares at a valuation of about $1.4 trillion, with the IPO price expected to be $70 to $78 per share. AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC will invest a combined $26 billion at the IPO price.
Trump backs Jensen Huang's return to Nvidia as CEO as US tightens AI chip exports the same day
Trump said he supports Jensen Huang returning to Nvidia as CEO, saying a large group of people wants him back and that no one has proven that Alice Huang, who was accused of being a spy, is indeed a spy. Huang resigned last Thursday over an extramarital affair report, ending a 33-year tenure; an open letter signed by Musk and other tech figures has gathered more than 165,000 signatures demanding that the board bring him back during an independent investigation. Nvidia's board reportedly has not considered reinstatement, and analyst Dan Ives sees no possibility of a return; on the same day, the US announced that companies in Singapore, Thailand, Malaysia, Indonesia, and Brazil must apply for export licenses for the most advanced chips. Nvidia shares closed down 4.1% on Monday, wiping out about $160 billion in market value to $3.76 trillion.
FBI warns open-source AI hacking tool Artex can breach networks in under three hours
The FBI issued a security advisory saying hackers are using open-source AI penetration-testing tools such as Artex to launch autonomous cyberattacks, with some attacks going from system targeting to actual intrusion in under three hours. Artex can connect to frontier large models from Anthropic, OpenAI, and others as well as open-source models, quickly scanning networks, using weak passwords and leaked credentials to gain a foothold, and then spreading laterally to steal sensitive information such as databases and email. Since its public debut in July, the tool has been used to attack multiple South Korean financial institutions, leaking the personal data of at least 68,000 people, and was also used to breach ConnectWise and access the Microsoft 365 environments of at least 2,250 customers; the developer has since made it closed-source, but old open-source versions are still circulating.
TypeSafe AI's Jev Popularizes Decision Models, Cloudflare and Other Companies Follow Suit
TypeSafe AI released Jev, a decision model billed as hallucination-free, and Cloudflare, Amazon, OpenAI, and Microsoft subsequently launched similar models. Jev claims end-to-end latency of 70 to 500 milliseconds and is priced at $0.042 per million input tokens with free output; a Carnegie Mellon University preprint test showed a median latency of 150 milliseconds, compared with 1.89 seconds for GPT-6 Astra. Researchers noted that decision models eliminate output hallucinations but can still make errors in mathematics, logic, and professional domains such as medicine, finance, and law; a hybrid approach that sends low-confidence results to an LLM for review slightly outperformed GPT-6 alone in tests, at 41% of the latter's cost.
Research papers
OpenAI Publishes Science Blog: GPT-6 Astra Assists Physicists in Proposing a Candidate Lattice Definition of the Standard Model
OpenAI published the first guest post on its science blog, in which Nathaniel Craig, head of the theoretical particle physics group at the University of California, Santa Barbara, and a visiting scientist at OpenAI, described his collaboration with GPT-6 Astra to propose a candidate lattice definition of the Standard Model. In a preprint completed in October and prepared with the help of GPT-6 Astra, Craig progressively mapped gauge fields on a raw lattice to finer lattices to compensate for phases; Astra proposed the mapping method, demonstrated that the comparison would converge to a definite answer, and led the establishment of the detailed proof. Lattice field theory experts Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed the proof and raised key questions; the scheme is currently more of an answer in principle, and Craig is investigating whether it satisfies reflection positivity at finite lattice spacing.
Briefs
- Mirendil Completes $1 Billion Series A, Bringing Total Funding to $1.2 BillionZhixiang Dongxi
- How the Claude Code Team Uses Verification Loops to Make Agents Self-Check CodeAnthropic: Claude.dev Developer Blog
- Nvidia Releases Robot Foundation Model GR00T 2, Plans to Open-Source It in DecemberHuxiu
- Baidu Q3 2026: AI Business Accounts for 53% of Core RevenueBaidu
- Berkeley RDI releases CyberArena: a cybersecurity benchmark where two AI agents attack and defend against each otherBerkeley RDI Blog
- OpenAI conducts embedded alignment evaluations with four institutions including Apollo and METROpenAI
- Microsoft Research and others propose TokenEconomizer: enabling coding agents to learn from historical sessions to reduce token costsMicrosoft Research
- Liquid AI releases d1.2 and d1.2-ultra omni-modal decision modelsLiquid AI Model and Engineering Blog
- Paul Christiano: superintelligence capable of overthrowing governments could emerge within a year; labs should first pause algorithmic progressDwarkesh Patel Podcast & Blog
- ARC Prize announces Inkling 2 Small's results on three ARC-AGI benchmarksARC Prize
AI News