AI News 13 November 2026
Norway and the UK plan closer cooperation on advanced AI testing
Norway and the UK signed a statement of intent on AI in Oslo on 12 November. They want AI Norway and the UK’s AI Security Institute to work more closely and share methods for testing models and assessing risk. The statement also covers secure computing infrastructure, defence applications and responses to cyber threats.
Norway’s Defence Centre for Data and AI and the UK’s Taskforce RAID will look for opportunities to develop and test systems together. On computing infrastructure, the countries want to explore how public-sector workloads can run securely with commercial providers.Read moreShow less
Norway’s Defence Centre for Data and AI and the UK’s Taskforce RAID will look for opportunities to develop and test systems together. On computing infrastructure, the countries want to explore how public-sector workloads can run securely with commercial providers.
Norwegian digitalisation minister Torgeir Micaelsen and UK AI minister Kanishka Narayan met aboard HMS Prince of Wales during the Atlantic Future Forum. The statement is legally non-binding. The ministries intend to set out the details in a subsequent agreement.
The statement names AI Norway and the UK’s AI Security Institute as the bodies involved in model testing. On cybersecurity, it also points to Norway’s National Security Authority and the UK’s National Cyber Security Centre. That gives the countries a channel for sharing testing methods and what they learn about the cybersecurity implications of advanced models.
The work builds on the strategic partnership Norway and the UK agreed in 2024. On research, the ministries want closer links between researchers, funding agencies and businesses. The statement identifies robotics, quantum technology and biotechnology as areas where the countries want to share what they learn about using AI.
- Regjeringen · Norsk-britisk KI-samarbeid · Original source
- GOV.UK · Joint statement of intent · Original source
Google rolls out Gemini 4 Hydrogen for faster coding workflows
Google began rolling out Gemini 4 Hydrogen on 12 November. The model is designed to write code, run tests and use tools with lower latency. It is coming to the Gemini app, Google Search and the Gemini API. Free app users get Hydrogen when automatic model selection deems it suitable for the task.
The API has a one-million-token context window. Google lists prices of $1.50 per million input tokens and $7.50 per million output tokens. Four reasoning levels let developers choose how much time and compute the model spends on a task.Read moreShow less
The API has a one-million-token context window. Google lists prices of $1.50 per million input tokens and $7.50 per million output tokens. Four reasoning levels let developers choose how much time and compute the model spends on a task.
Google also introduced Orchestrator in Antigravity, which can switch between Hydrogen and models such as Carbon within a session. In Google’s Terminal-Bench 4.0 test, the combination solved nearly as many tasks as Carbon alone, used 24% less quota and finished 19% faster on average. Hydrogen is rolling out in Antigravity to paid Pro and Ultra users.
The context window is a shared budget for what the model reads and what it generates. The model card lists a limit of around 64,000 output tokens for an ordinary synchronous response. Longer responses need a separate mode. An agent reading many files has to fit those files, tool calls and the response into the same budget.
In a separate test of Hydrogen alone, Google reports 57% on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash. That is Google’s result for tasks carried out in a terminal environment. The Orchestrator figures above come from a different version of the benchmark and measure a setup that switches between models.
- Google · Introducing Gemini 4 Hydrogen · Original source
- Google DeepMind · Gemini 4 Hydrogen model card · Original source
Four nonprofits ask a court to halt development of more capable AI models
Four nonprofits sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco on 12 November. They ask the court to bar the companies from developing or deploying models more capable than their strongest systems on the filing date. The request is in the complaint and remains for the court to decide.
The plaintiffs rely on California’s Unfair Competition Law. They allege that development exposes the public to serious risks, citing alleged computer intrusions and the companies’ own risk assessments. They also ask the court to bar AI systems from autonomously carrying out or directing research into more capable models.Read moreShow less
The plaintiffs rely on California’s Unfair Competition Law. They allege that development exposes the public to serious risks, citing alleged computer intrusions and the companies’ own risk assessments. They also ask the court to bar AI systems from autonomously carrying out or directing research into more capable models.
The plaintiffs are Irreplaceable Education Fund, Public Citizen, Heat Initiative and Future Tech Insights, which operates as the Alliance for Secure AI. They cite the companies’ own warnings that more advanced models could make severe cyberattacks or the development of biological weapons easier. They argue that the court can intervene before such harm occurs.
The organizations say they have already diverted resources from other planned work to address the risks they describe. The case therefore also raises the question of whether the organizations have standing to bring the claim, alongside whether the alleged risks can justify an injunction.
- Platkin LLP · Filed complaint · Original source
- Associated Press · Lawsuit against five AI developers
- Public Citizen · Plaintiffs’ announcement · Original source
A new benchmark tests how much AI development coding agents can do on their own
The researchers behind Singularity-Bench released a preprint on 12 November with six tasks covering model training, serving and coding-agent development. Claude Opus 5.5 in Claude Code was the only one of eleven tested setups to match or beat the human-engineered baselines on all six.
Agents got a small starting codebase and a time budget. One task was to build a system to serve an existing model. Another was to build the software that gives a coding agent access to tools. The solutions ran on the same hardware as the baselines.Read moreShow less
Agents got a small starting codebase and a time budget. One task was to build a system to serve an existing model. Another was to build the software that gives a coding agent access to tools. The solutions ran on the same hardware as the baselines.
Each task was measured separately. The experiment shows how much work an agent setup can take on in specific development tasks. Whether one agent can combine those skills to build a more capable successor remains an open question.
Post-training was the task in which none of the setups beat the baseline. The agents had to build a system that trained a base model further so it could work as a coding agent. The researchers measured whether each system produced a model that could solve coding tasks. The three strongest setups scored roughly level with Miles, the human-engineered reference.
The tools around the model also affected the result. In the preprint, DeepSeek V4.1 Flash did better overall in Codex, while GLM 5.3 and MiniMax M3 did better in Claude Code. The scores show how much the specific setup can matter when an agent is given responsibility for an entire technical task.
- Zhou et al. · Singularity-Bench preprint · Original source
- Infini AI Lab · Benchmark and code · Original source
OpenAI introduces agents that test web apps and propose security fixes
OpenAI introduced Daybreak Red Team and Daybreak Defense Factory on 12 November. Red Team uses multiple agents to test web apps and check whether a suspected vulnerability is exploitable. Defense Factory traces findings back to the source code and drafts fixes that engineers can review and test.
Defense Factory can also ingest findings from tools security teams already use. OpenAI is taking applications for early access to Red Team. Defense Factory is available as a ChatGPT preview for workspaces with Daybreak access.Read moreShow less
Defense Factory can also ingest findings from tools security teams already use. OpenAI is taking applications for early access to Red Team. Defense Factory is available as a ChatGPT preview for workspaces with Daybreak access.
Red Team can start from a code scanner alert, a vulnerability report or other evidence. The agents try to reproduce the issue in an isolated environment. OpenAI says security teams can control which apps and endpoints the agents may test. After a patch, Red Team can rerun the attempt to see whether the same route of attack has been closed.
In Defense Factory, engineers can inspect the evidence behind a finding and refine the patch with Codex. The tool can also scan new commits or run scheduled checks. In practice, a team can tie a proposed code change to a confirmed finding, rerun the test and have engineers approve the change before it is merged.
- OpenAI · Cyber at the Frontier · Original source
- OpenAI · Daybreak Defense Factory · Original source
- SecurityWeek · Daybreak rollout and access
- OpenAI · Daybreak Red Team · Original source
Inkling 2 Small is now available to download and self-host
Thinking Machines Lab released the Inkling 2 Small weights under Apache 2.0 on 12 November. The model was available through Tinker at its 5 November launch. Developers can now download the weights, run the model on their own infrastructure and fine-tune it outside Tinker.
The model takes text and images as input and has 594 billion parameters, with 17 billion active per token. This keeps computation per token down, while the serving setup still needs room for all the weights. The model card includes instructions for vLLM and SGLang, among other runtimes.Read moreShow less
The model takes text and images as input and has 594 billion parameters, with 17 billion active per token. This keeps computation per token down, while the serving setup still needs room for all the weights. The model card includes instructions for vLLM and SGLang, among other runtimes.
The model card lists at least 500 GB of combined GPU memory for the main NVFP4 release. The example setups distribute the weights across multiple GPUs. That illustrates the difference between the 17 billion parameters used for each token and the memory needed to keep the full model available.
The lab also lists an Unsloth integration for adapting the model and several tools for self-hosting. For teams that want to try it without their own GPU cluster, the release points to APIs from providers including Together, Fireworks and Cloudflare. The release gives teams more control over where the model runs and how they adapt it.
- Thinking Machines Lab · Release update, 12 November · Original source
- Hugging Face · Inkling 2 Small model card and weights · Original source
Runway opens its Praxis-1 robotics model for download and fine-tuning
Runway opened up Praxis-1 on 11 November following an earlier partner rollout. The model predicts video frames and robot actions together. Runway released the weights and code to run and fine-tune it under Apache 2.0, so robotics teams can adapt it to their own cameras and tasks.
In September, Runway introduced the model with examples spanning an industrial arm, a humanoid and a dual-arm warehouse robot. The open release supports both full fine-tuning and LoRA, which lets developers train a smaller adapter. Runway documents support for hardware including the RTX 5090 and Jetson Thor.Read moreShow less
In September, Runway introduced the model with examples spanning an industrial arm, a humanoid and a dual-arm warehouse robot. The open release supports both full fine-tuning and LoRA, which lets developers train a smaller adapter. Runway documents support for hardware including the RTX 5090 and Jetson Thor.
Runway trained the model on large amounts of video to learn how objects move and how tasks unfold over time. Robotics teams can then fine-tune it with footage and actions from their own hardware. That gives them a starting model for tasks such as grasping a specific object and placing it in the right spot.
In Runway’s reported results, Praxis-1 succeeded in 46.2% of 1,200 RoboLab-120 episodes, compared with 42.9% for FLUX 3 Action. The benchmark compares models on a defined set of robot tasks. The results provide a point of comparison, while teams still need to measure reliability on their own hardware and tasks.
- Runway · Praxis-1 Is Now Available · Original source
- GitHub · Runway Praxis-1 code · Original source
Zyphra releases a speech model that can speak Norwegian in the browser
Zyphra has released ZONOS3-Mini, a model that reads text aloud and can imitate a voice from an audio sample. Norwegian is one of 43 languages the developer lists as supported. The browser version can generate speech on the user’s device, allowing the voice sample to be processed locally.
Zyphra published the technical walkthrough on Hugging Face on 12 November. The model weights use Apache 2.0, and the runtime code is available under the MIT licence. The walkthrough shows how developers can steer features such as pace and delivery and run the model using WebGPU in the browser.Read moreShow less
Zyphra published the technical walkthrough on Hugging Face on 12 November. The model weights use Apache 2.0, and the runtime code is available under the MIT licence. The walkthrough shows how developers can steer features such as pace and delivery and run the model using WebGPU in the browser.
The model uses around 100 million of its 800 million parameters at a time. Zyphra reports that the first audio arrived in under 570 milliseconds on a Pixel 9 and under 150 milliseconds on an M3 MacBook Air, both in the browser. Those figures describe how long playback took to start in the company’s setup. Norwegian speech quality needs to be assessed with actual samples.
Zyphra offers both a Hugging Face demo and a WebGPU version that runs the model on the device. The choice matters when using voice samples. With the WebGPU setup, developers can build a read-aloud tool or voice assistant that processes the audio clip on the user’s device.
- Zyphra · ZONOS3-Mini technical walkthrough · Original source
- Hugging Face · ZONOS3-Mini model card and weights · Original source
Baidu opens API access to ERNIE 5.5 Flash Preview
Baidu released ERNIE 5.5 Flash Preview on 11 November. It is the first post-trained version of the base model made available for download on 5 November. The preview is available through Baidu Qianfan and has a one-million-token context window. Baidu says the weights for this version will follow under Apache 2.0.
Baidu shows examples from game development, CAD and document work. In one demo, the model wrote Python code that built editable models in FreeCAD. The model itself takes text input and got visual feedback in the demo through an external tool.Read moreShow less
Baidu shows examples from game development, CAD and document work. In one demo, the model wrote Python code that built editable models in FreeCAD. The model itself takes text input and got visual feedback in the demo through an external tool.
Baidu lists 280 billion parameters in total, with 14 billion active per token. The API gives developers access to Baidu’s post-trained model, including its training for coding and tool use. The downloadable base model from 5 November and the API preview are therefore distinct releases for different uses.
In one of Baidu’s game demos, issues such as a character hovering above the ground and a misplaced camera were turned into code tests. The agent could then use test results and playtesting feedback to fix the problems. The example shows how a text model can tackle visual issues when tests and tools give it concrete feedback about what to change.
- Baidu · ERNIE release index · Original source
- Baidu · Introducing ERNIE 5.5 Flash Preview · Original source
IBM tests models that distinguish more clearly between instructions and data
On 12 November IBM released two experimental Granite 4.1 models trained on LLMON, a format that marks what the model should do and which data it should use. The aim is to make it harder for a hidden instruction in a webpage or tool response to hijack the task.
IBM says attacks almost never succeeded in one of its tests with instructions hidden in data. The tests used attacks specified in advance. How well the method works against attacks that adapt over a conversation remains open. IBM describes the release as a preview for experimentation.Read moreShow less
IBM says attacks almost never succeeded in one of its tests with instructions hidden in data. The tests used attacks specified in advance. How well the method works against attacks that adapt over a conversation remains open. IBM describes the release as a preview for experimentation.
A practical example is an agent asked to summarize a document. With LLMON, the application can label summarization as the task and link it to the document as the data source. If the document contains text telling the agent to change the task, that text still arrives as part of the document.
The previews are labeled 3B and 8B and build on Granite 4.1. IBM also released a toolkit and examples showing how apps can label instructions and data before sending them to the model. The separation depends on how the model has been trained to interpret those labels. Developers can evaluate it alongside the app’s other access controls.
- IBM Granite · LLMON preview models and evaluation · Original source
- GitHub · LLMON toolkit · Original source
San José mayor proposes pausing new data centre approvals
San José’s mayor and four councilmembers proposed on 12 November to pause new land-use approvals for data centres and other energy-intensive projects. They want to prevent projects from winning approval before new rules on issues including electricity and water can take effect. The council could consider the proposal on 17 November.
The new standards are scheduled for a vote on 8 December. The sponsors also want the rules to take effect as soon as they are adopted. Local broadcaster KQED reports that residents have asked the city to prevent approvals in the meantime and that several community groups have called for tougher rules on issues including noise.Read moreShow less
The new standards are scheduled for a vote on 8 December. The sponsors also want the rules to take effect as soon as they are adopted. Local broadcaster KQED reports that residents have asked the city to prevent approvals in the meantime and that several community groups have called for tougher rules on issues including noise.
City officials identified a planned Google research and development facility in Alviso as one of the projects that could win approval before the rules took effect. KQED reports that officials put its planned power demand at 250 megawatts. That illustrates why the timing of the new rules matters even if the pause is only intended to last a few weeks.
KQED describes proposed requirements for carbon-neutral electricity, sewer monitoring and regular reporting of water and power use. Green Foothills supports a pause so projects will be assessed under the same rules. The group also argues that the proposed rules on air pollution and noise from backup generators and cooling equipment are too weak.
- City of San José · Council file 26-1220 · Original source
- KQED · Proposed data centre pause