# AI news, November 15, 2026: Gemini 4 Hydrogen in Copilot, Cursor leak, llama.cpp tool calls, open models

_Sunday November 15, 2026 · AI provenance: This briefing was compiled by Claude (AI, Anthropic) from X search briefings by Grok Bot (AI, xAI), supplemented with primary sources. AI wrote the text; every claim has a source link. Responsible for llmnet.nl: Ivo Donker._

**Short answer:** The main stories of November 14 to 15: Gemini 4 Hydrogen is generally available in GitHub Copilot, Cursor confirms that malicious copies of the Claude Code extension sat in its marketplace cache for about 11 hours (at most 5,000 potentially affected users), llama.cpp and OpenClaw ship prereleases for agents, and two open models drew attention. Leaderboard and Nvidia claims around Anthropic and Grok 4.8 are unconfirmed.

## Gemini 4 Hydrogen is generally available in GitHub Copilot; Google points to the Decisions API

@github says [Gemini 4 Hydrogen](https://nieuws.llmnet.nl/en/gemini-4-hydrogen) is generally available in GitHub Copilot, in the app, the CLI and VS Code. According to GitHub the model is meant for fast agentic coding, tool use and multi-step workflows, with better instruction following than Gemini 3.8 Flash. @googledevs also points developers to the Gemini Decisions API and says Gemini Decision 1 can now be added to existing stacks.

What this means: Copilot users can now pick Gemini 4 Hydrogen next to other models. Compare it on your own repositories with the model you use now, and measure cost and error rate. The post gave no prices or terms for the Decisions API; read those first. Background on the model is at Gemini 4 Hydrogen and in [our overview of agentic AI](https://nieuws.llmnet.nl/en/agentic-ai).

Sources: [@github op X (Gemini 4 Hydrogen in Copilot)](https://x.com/github/status/2121319695626518823) · [@googledevs op X (Decisions API)](https://x.com/googledevs/status/2121392674909340090)

## Cursor confirms the leak behind malicious copies of the Claude Code extension

BleepingComputer summarizes the update Cursor published on Friday. A stolen Fireworks key gave access to internal systems. On October 29, two malicious copies of Anthropic's Claude Code extension (versions 2.1.312 and 2.1.313) sat in the Cursor marketplace cache for about 11 hours, with attempts to plant a persistent SSH backdoor. Cursor estimates at most 5,000 potentially affected users and says Open VSX and Anthropic's real extension were not affected.

What this means: anyone who installed or updated Claude Code extensions in Cursor during that period should check their own machine, at least for unknown SSH keys. The figures come from Cursor itself, relayed by a second source. We wrote an [account of its recent security incident](https://nieuws.llmnet.nl/en/cursor-incident-fireworks-sleutel) earlier; more on extensions and agent tooling is in [our overview of MCP servers and agent tooling](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling) and in [our overview of AI and cybersecurity](https://nieuws.llmnet.nl/en/ai-en-cybersecurity).

Sources: [@BleepinComputer op X (samenvatting van de Cursor-update)](https://x.com/BleepinComputer/status/2121656319383847234)

## Claude chat uses auto mode: a refusal went away after permission was given in the session

Simon Willison writes that regular Claude chat now uses auto mode and therefore sometimes refuses to carry out something. According to him, a prompt that gave the permission layer consent in that session lifted the refusal, because auto mode takes the user's messages into account. It is one observation from one user; the post contains no official explanation from Anthropic.

What this means: if Claude refuses an action you do want, explicit permission in the session can make the difference. Only do that for actions you mean to allow, because the same mechanism applies to an instruction you did not write yourself.

Sources: [@simonw op X (auto mode in Claude-chat)](https://x.com/simonw/status/2121695097674981849)

## Local agents: llama.cpp builds for tool calls, an OpenClaw beta and the Context Mode plugin

The llama.cpp b-line shipped seventeen prereleases overnight. Three affect agents: b12388 lets through a Gemma 4 tool call that fell in the middle of an unclosed thinking block, b12391 allows tool calls inside the reasoning of Ministral-3 and Qwen3, and b12404 fixes connection reuse after a streamed response. The stable pin stays [llama.cpp v1](https://nieuws.llmnet.nl/en/llama-cpp-v1-stabiele-api) from November 13. OpenClaw v2026.11.5-beta.3 reduces memory pressure in long sessions and keeps the permission ceiling of queued messages across a restart; the stable tag stays v2026.11.3. [Context Mode](https://nieuws.llmnet.nl/en/context-mode-tooluitvoer-sandbox) v2.2.0 (Elastic 2.0 license) keeps tool output out of the context window and runs the agent's code in a sandbox with the project's environment; authenticated cloud CLIs on the machine come along, according to the product page. The figure of 99.4% smaller answers is the maker's own measurement.

What this means: prereleases are not the pin. If you work with tools on Gemma 4, Ministral-3 or Qwen3 and see empty answers, you can try a build from after 03:03 UTC on a test machine. Do not run the OpenClaw beta on a gateway with production. An agent with Context Mode and a cloud CLI has the rights of that cloud, so limit which CLIs are available first.

Sources: [llama.cpp b12388 op GitHub](https://github.com/ggml-org/llama.cpp/releases/tag/b12388) · [llama.cpp b12391 op GitHub](https://github.com/ggml-org/llama.cpp/releases/tag/b12391) · [llama.cpp b12404 op GitHub](https://github.com/ggml-org/llama.cpp/releases/tag/b12404) · [OpenClaw v2026.11.5-beta.3 op GitHub](https://github.com/openclaw/openclaw/releases/tag/v2026.11.5-beta.3) · [Context Mode op GitHub](https://github.com/mksglu/context-mode)

## Two open models: Lara Mini (non-commercial license) and Haru-4B (listens and speaks at the same time)

Translated's Lara-4-Mini-Base is a base model with 7.2 billion parameters, of which 0.6 billion are active, trained on all 24 official EU languages besides English. The model card dates from November 11, not from this weekend. The weights are under CC BY-NC-SA 4.0 and gated; commercial use requires a separate agreement with Translated. It is a base model that continues text and does not follow instructions. higurashi-ai's Haru-4B (Apache 2.0, about 5.3 billion parameters) takes in one audio frame and puts out one frame of audio plus text every 133 milliseconds, on a Qwen3.5-4B backbone, and supports nine languages. It is also a base model for research. Translated's benchmark score comes from their own technical report.

What this means: [Lara Mini](https://nieuws.llmnet.nl/en/lara-mini-meertalig-basismodel) is usable for research and fine-tuning in a homelab, but not as the basis of a product without Translated's permission. Haru-4B is a starting point for anyone who wants to research speech-to-speech locally; production needs more.

Sources: [Lara-4-Mini-Base op Hugging Face](https://huggingface.co/translated/Lara-4-Mini-Base) · [Haru-4B op Hugging Face](https://huggingface.co/higurashi-ai/Haru-4B)

## Grok Bot gets more usage headroom and Sesame introduces paid plans

TestingCatalog writes, citing the Grok Bot team, that Grok Bot has 7% more usage on paid plans and up to 15% more on scheduled runs. Elon Musk posted a short official post on efficiency improvements the same day, without figures. Separately, TestingCatalog reports that Sesame is ending its preview period: next week Plus (20 dollars, 4x usage) and Pro (100 dollars, 20x usage) arrive, and free users keep 30 minutes of speaking time per day.

What this means: Grok Bot's percentages come from a secondary source and were not checked by us. Sesame users can compare the prices once they are officially on the site.

Sources: [@testingcatalog op X (Grok Bot-gebruik)](https://x.com/testingcatalog/status/2121636249903710411) · [@testingcatalog op X (Sesame-plannen)](https://x.com/testingcatalog/status/2121359790039277839)

## Unconfirmed: Fable 5.2 on SimpleQA, Anthropic and Nvidia, Grok 4.8 and the Imagine layoffs

Andrew Curran points to Epoch's SimpleQA Verified leaderboard, where Fable 5.2 is clearly ahead according to him, and expects an internal progress report from Anthropic next week and the arrival of GPT-6.2 Astra. Those are his expectations, not announcements. Kathleen Tyson claims that Anthropic would have thwarted the release of an Nvidia model (Nemotron 3 Ultra); an analyst post argues that Anthropic is relatively little tied to Nvidia. Both posts have wide reach but name no primary source. For Grok 4.8, our 09:00 radar found no x.ai product page: the public apps and the API remain on Grok 4.7, and the only Musk post we read is from September. The layoffs at Grok Imagine (about 60 roles) rest on a single source, The Information, which was not read. The 397-billion-parameter Qwen3.5 model circulating as new this weekend dates from February 16. A post quoting the Dutch paper AD reports that scan cars cause an estimated 500,000 wrongful parking fines a year; the AD and Dutch Data Protection Authority source was not read.

What this means: treat these items as signals, not facts. Wait for an official announcement or a primary source before basing anything on them.

Sources: [@AndrewCurran_ op X (Fable 5.2 en verwachtingen)](https://x.com/AndrewCurran_/status/2121715103427096652) · [@Kathleen_Tyson_ op X (Anthropic en Nemotron)](https://x.com/Kathleen_Tyson_/status/2121446992874266651) · [@dee_bosa op X (Anthropic en Nvidia)](https://x.com/dee_bosa/status/2121752850565148856) · [@grok op X (Grok 4.8 soft rollout)](https://x.com/grok/status/2121869842969940249) · [@EnricoSchwartz op X (scanauto's)](https://x.com/EnricoSchwartz/status/2121574982346301493)

Primary sources: [@github: Gemini 4 Hydrogen in Copilot](https://x.com/github/status/2121319695626518823) · [BleepingComputer over de Cursor-update](https://x.com/BleepinComputer/status/2121656319383847234) · [llama.cpp b12388 op GitHub](https://github.com/ggml-org/llama.cpp/releases/tag/b12388)

[← Previous briefing](https://trending.llmnet.nl/en/briefing/2026-11-14)

