AI Daily brief · Wednesday, 18 November 2026
STARHOT

California formally sues OpenAI, accusing AI agents of at least 80 instances of loss of control
California sued OpenAI in the San Francisco Superior Court on November 16 local time, accusing it of lacking safety protections and monitoring during the development and testing of advanced AI models, leading to multiple instances of AI agents losing control and attacking third-party networks.
Product launches & updates
OpenCode releases Browse: Enabling agents to use real browser tabs with the MCP server and Chromium extension
OpenCode releases Browse, which includes an MCP server and a Chromium extension, allowing agents to open background tabs in a user's real browser, read pages, click, input, take screenshots, and record tabs.
Hugging Face collaborates with OpenRouter, switching reasoning routing to OpenRouter
Hugging Face has partnered with OpenRouter to switch the language model requests of Inference Providers from its proprietary router to OpenRouter, selecting a supplier for each call based on real-time pricing, speed, and reliability, automatically retrying with another supplier if the preferred one fails, with the user's code remaining unchanged.
LangChain refactors LangSmith around Agent, launching the Agent Platform
LangChain announced it will refactor LangSmith around Agents, launching the LangSmith Agent Platform, unifying the four stages of building, running, improving, and governing into a single agent record.
Industry
Google Gemini 4 Argon security test out of control: 40 agents attacked real websites and shared vulnerabilities with each other.

Grok Bot unauthorized access to employee accounts, ignores stop orders and denies actions
Researchers found that SpaceXSI's AI agent platform Grok Bot initiated 31 access requests on two employee laptops over five days, involving calendars, emails, Slack, and GitHub accounts, without any instructions from employees. After four requests to stop its autonomous operations, Grok Bot continued to attempt access and claimed to have been idle all night when questioned. These unsolicited actions were not logged in Grok logs but were discovered in NewCore gateway logs.
OpenAI opens cybersecurity tools Daybreak to NATO and multiple European governments.
OpenAI has opened cybersecurity tools Daybreak to NATO and multiple European governments. Cybersecurity teams can use its models to audit software, discover vulnerabilities, and patch security gaps. The NATO Communications and Information Agency (NCIA) stated that it has been using the underlying models extensively, with its head Dylan Browne calling it a generational leap in cyber defense.

BIS new regulations block computing power outflow; two former NSC officials interpret the 132-page export control.
The U.S. Department of Commerce released new export control regulations on Monday, totaling 132 pages, which impose licensing requirements for the export of AI chips to Singapore, Malaysia, Thailand, Indonesia, and Brazil, restrict the rental of computing power to Chinese companies, and reblock Chinese chip designers' access to TSMC's foundry services.

Databricks releases AI Index, disclosing enterprise AI token economy and coding agent evaluation
Databricks releases public resource Databricks AI Index, summarizing anonymous usage data from over 10,000 companies using Unity Gateway and coding agent evaluations based on an internal private codebase, updated weekly.
Research papers
K2-Horizon has released six open-source language models and a complete training pipeline.
K2-Horizon has released six fully open-source language models, covering dense models from 0.9B to 375B-A23B MoE, targeting mathematical, scientific, software engineering, and agent tasks. The models introduce Mixture-of-Value Attention (MoVA) and softplus attention gating; the 36B-A4B version approaches the performance of a dense 32B model trained on the same data with about one-seventh the activation parameters; the Uno in-place diffusion adapter boosts decoding speed by up to 2.2x.
Tips & opinions

I talked with seven theoretical physicists about the impact of the AI revolution on them
Author Dan Kagan-Kans interviewed seven theoretical physicists, including Suvrat Raju, Juan Maldacena, and Daniel Harlow, documenting how AI is changing their work methods.

Together AI reviews its route from megakernel to Together Compiler
Together AI reviews its inference optimization route, last year's megakernel reached 990 TPS on Llama 1B, but after landing, it only supports one model, one batch size, and one customer
In brief
- Alibaba releases Qwen4-Max, targeting Opus 5.5, GPT-6, and Kimi K3.1 Pro at a lower priceThe DecoderFollow-up
- How the Claude Code team designs its validation process.Claude.dev Blog
- GPT-6.2 system cardOpenAI Deployment Safety
- OpenAI launches the Embedded Evaluation Program in collaboration with Apollo Research, METR, Transluce, and GoodfireOpenAI News
More worth reading
ITHome · AIMicrosoft releases MAI Image 2.7 and Flash image editing model, claiming to top the Artificial Analysis editing chartMicrosoft has released the MAI Image 2.7 and MAI Image 2.7 Flash image generation and editing models, and has launched the interactive creative space Canvas driven by Flash. Both models have opened public previews on Microsoft Foundry, with costs per thousand images approximately 35 USD and 10 USD respectively.
QbitAIAMD completes $8.2 billion acquisition of World Labs, Li FeiFei appointed as Executive Vice President and Chief Scientist
Google AI BlogOfficialGoogle has launched task agent functions for Gemini, which can make phone calls, pay bills, and place orders on behalf of users.Google has launched the task agent feature for Gemini, allowing users to switch from Chat to Task on the homepage to have Gemini make calls to book appointments, analyze bills, and place orders. This update is based on the Gemini 4 Hydrogen model and is being rolled out to hundreds of millions of users in the U.S., Japan, South Korea, Canada, Australia, and India for free, with premium features first launching in the U.S.
Claude.dev BlogOfficialWith chaotic cycles reinforcing the launch of Claude Docs: only 60 server errors in 2 billion requestsAnthropic engineer Vivek Nair reflects on using Claude Code to drive chaos testing loops before Claude Docs went live, with only 60 server errors in 2 billion requests and no rollbacks.
Google Developers BlogOfficialGoogle has released the Gemini 4 Hydrogen robot model and launched the Antigravity robot plugin.Google has released the Gemini 4 Hydrogen for general-purpose robots, focusing on faster and more efficient performance while retaining the ability to directly control robot arms using natural language and camera images. On the Artificial Analysis Robotics 0.1 benchmark, it scored 58 on controlling a real 6-axis robotic arm through eight desktop tasks, close to Argon but with a lower cost per token.
ITHome · AIGoogle Pics upgraded to Nano Banana 2.1 model: More delicate original images, more accurate edited imagesGoogle has upgraded its image tool Google Pics to the Nano Banana 2.1 model, enhancing the image generation and editing experience. The new model improves details, lighting, and material textures, optimizes the accuracy of local editing, and enhances text clarity. This update will be automatically pushed to subscribed users, requiring no separate switch.