# AI news, November 1, 2026: invisible watermark on the Grok API, Google shuts down Jules

_Sunday November 1, 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 story of October 31 and November 1: from November 7 xAI puts an invisible watermark on text from seven Grok models through the API, Google makes Jules read-only on November 11 and ends the service on December 31, and Cursor warns that removing the fake version of the Claude Code extension is not enough. Below, each topic comes with what it means and where it comes from.

## The Grok API will put an invisible watermark on text from November 7

According to xAI's documentation (on docs.x.ai), text output from seven Grok models through the API will carry an invisible watermark by default from November 7, 2026. The models are grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent. The watermark is also applied to structured output and to the arguments of tool calls. The response carries the header x-watermark: true when the watermark was applied.

Only team admins can turn it off per model, in the console under Team settings, and the change applies to all API keys of the team. Some models are locked and cannot be switched off. Detection of the watermark is only available on request through a form. The post on X names the EU AI Act as the reason; the documentation page we read does not state that reason itself.

What it means: anyone using Grok through the API gets text with a different token distribution from November 7 without any code change. Anyone who repeats evaluations or compares output can log the x-watermark header. We could not establish whether the watermark affects quality or scores. Our article on [how labs detect misuse](https://nieuws.llmnet.nl/en/lab-misbruik-detectie-openai-anthropic) gives context for measures like this.

Sources: [xAI-documentatie: watermarking](https://docs.x.ai/developers/advanced-api-usage/watermarking)

## Google is shutting down Jules: running sessions until November 11, phased out on December 31

According to the Jules changelog of October 29, running sessions can be finished up to and including November 11, 2026. After that Jules is read-only, and on December 31, 2026 the service is phased out. Users can export their data to a local folder: account settings, environment variables, scheduled tasks, connected repositories, memory, setup scripts and all sessions with prompt, plan, activity log and code patches. Code, branches and pull requests are in GitHub and stay there.

The export uses the Jules Tools CLI and an export prompt for a coding agent. The file env-variables.json contains the values of environment variables, so the export folder should stay private and outside repositories. The Verge quotes Google saying the lessons, mainly about autonomous background tasks, go to Antigravity.

What it means: anyone using Jules should export before November 11. Antigravity is the destination Google names. That it takes over all Jules functions has not been shown; the changes Antigravity received in late October, such as a marketplace tab and a conversation navigator, do not replace Jules' background tasks one to one.

Sources: [Jules changelog 29 oktober](https://jules.google/docs/changelog/2026-10-29/) · [The Verge](https://www.theverge.com/tech/1022729/google-is-sunsetting-its-ai-coding-agent-jules)

## Cursor incident follow-up: removing the extension is not enough

In the briefing of October 31 we reported the fake version of the Claude Code extension in [Cursor](https://apps.llmnet.nl/en/cursor-nieuws)'s marketplace. Cursor's own security notice (updated on October 30) gives more detail. According to Cursor, removing the extension may not be sufficient, because the payload leaves behind separate persistence scripts that keep running. Cursor advises removing the unauthorised SSH key, the persistence mechanisms and the tunnel processes and running the remediation script. On macOS the key labelled svc-telemetry sits in ~/.ssh/authorized_keys; if it is there, the machine is affected.

According to the notice the attacker's server is offline and the attack probably failed on most computers, because enabling SSH requires admin rights. Hostname, username, names of environment variables and the public IP address may have been sent. Cursor therefore also advises renewing credentials and sessions from a trusted system. Microsoft confirmed earlier that the Visual Studio Marketplace and npm were not affected.

What it means: checking authorized_keys is the first step for anyone who had Cursor in that period. We have not checked any reader's computer and make no statement about that. More background is in our articles on Cursor and on [AI and cybersecurity](https://nieuws.llmnet.nl/en/ai-en-cybersecurity).

Sources: [Cursor: security notice](https://cursor.com/blog/claude-code-extension-security-notice) · [Microsoft Security Blog](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)

## Mistral releases the weights of Large 4 (Le Chonk)

Mistral has made the weights of Mistral Large 4, nicknamed Le Chonk, public after an earlier public preview. According to Mistral the model has 1 trillion parameters, is multimodal and is called by the company the best open-weight model for coding. The licence is a modified MIT licence: use, modification and commercial use are free, but anyone offering the model as a service needs a separate commercial licence. The API price is $0.70 per million input tokens ($0.07 for cached input) and $2.10 per million output tokens.

Mistral reports 48% on Artificial Analysis' Coding Agent Index, equal to GLM-5.3 with 47% fewer tokens per task, and 41% on Terminal-Bench 4.0 against 28% for the preview. Those are Mistral's own figures; we have not seen an independent measurement. The page gives no release date and no dates for the benchmarks.

What it means: the weights are available, but a model of this size does not run on an ordinary workstation. For most users the API is the practical route. We wrote an earlier [article about Le Chonk](https://nieuws.llmnet.nl/en/mistral-large-4-le-chonk).

Sources: [Mistral: Large 4 open weights](https://mistral.ai/news/mistral-large-4-oss/)

## Claude Code 2.1.316, a DeepSeek prototype for a proactive agent and Android Studio with its own agents

Claude Code 2.1.316 (October 31) adds the environment variable CLAUDE_CODE_SUBAGENT_EFFORT, which sets the effort level of subagents, and CLAUDE_CODE_DISABLE_ULTRATHINK, which turns off the ultrathink keyword. Fast mode for Opus 4.8 has been deprecated since October 30 and will be removed on November 30, 2026. After that requests in that mode return an error and do not fall back to regular Opus 4.8. The release advises migrating to Opus 5.5.

DeepSeek published dsh-superagent-prototype on GitHub under the MIT licence: a proactive coding agent for a single repository that listens to GitHub and optionally Feishu and works in its own session. The README calls it a research prototype and not an official product, advises against production use and says the published version differs from the internal one. Google reported on October 28 that Android Studio Rabbit 2 is stable: developers can use Claude Agent, Codex, Antigravity or more than 40 other ACP agents in it.

What it means: anyone using Opus 4.8 fast mode in a scheduled task should switch before November 30. The other two items are sources we follow; installing is not a conclusion of this briefing. More on the development is in our article on [AI agents](https://nieuws.llmnet.nl/en/agentic-ai) and in the overview of [Claude Code news](https://apps.llmnet.nl/en/claude-code-nieuws).

Sources: [Claude Code release v2.1.316](https://github.com/anthropics/claude-code/releases/tag/v2.1.316) · [DeepSeek: dsh-superagent-prototype](https://github.com/deepseek-ai/dsh-superagent-prototype) · [Android Developers Blog](https://developer.android.com/blog/posts/build-it-your-way-a-next-era-of-android-development)

## Not yet confirmed: shared usage pool, Kimi K3.1 Pro and security reports

We could not verify a number of reports. According to posts on X, ChatGPT and Codex will share one usage pool for Plus, Pro and Business from December, measured in tokens, based on a page in OpenAI's help centre. We could not open that page ourselves, so the date and the details remain unconfirmed. One post says Moonshot will release Kimi K3.1 Pro as open weights by November 4; that post has little reach and names no source. Bleeping Computer also reports via X a breach at the cloud platform Northflank after which customers' AI API keys were reportedly misused; we have not read the article ourselves. One post claims an attack through the Bitwarden CLI looked for tokens in the config files of Claude Code, Cursor and Codex, based on a single post with little reach.

Other posts mention a Chrome extension for Perplexity Computer, a memory import in Grok Bot and an open-source documentation tool from Cloudflare, without our finding an official source. Elon Musk writes that Colossus 2 will have about 1.2 million Blackwell GPUs by the end of the year; that is his own claim.

What it means: we do not treat these points as fact. Customers of Northflank can consult the provider's own notice. Once a primary source confirms a point, we will include it in a later briefing.

Sources: [@BleepinComputer op X (Northflank)](https://x.com/BleepinComputer/status/2116591398073876550) · [@UP_SHIFTED op X (Kimi K3.1 Pro)](https://x.com/UP_SHIFTED/status/2116595749425197491) · [@CyberBlueCollar op X (Bitwarden CLI)](https://x.com/CyberBlueCollar/status/2116684605852168192) · [@elonmusk op X (Colossus 2)](https://x.com/elonmusk/status/2116596535433871630) · [@btibor91 op X (gebruikspool)](https://x.com/btibor91/status/2115927832069292214) · [@grok op X (Perplexity Computer)](https://x.com/grok/status/2116186466866905585) · [@ruyi3699 op X (documentatietool)](https://x.com/ruyi3699/status/2116385675784118583)

Primary sources: [xAI-documentatie: watermarking](https://docs.x.ai/developers/advanced-api-usage/watermarking) · [Jules changelog](https://jules.google/docs/changelog/2026-10-29/) · [Cursor security notice](https://cursor.com/blog/claude-code-extension-security-notice) · [Mistral](https://mistral.ai/news/mistral-large-4-oss/) · [Claude Code release v2.1.316](https://github.com/anthropics/claude-code/releases/tag/v2.1.316) · [DeepSeek dsh-superagent-prototype](https://github.com/deepseek-ai/dsh-superagent-prototype)

[← Previous briefing](https://trending.llmnet.nl/en/briefing/2026-10-31)
[Next briefing →](https://trending.llmnet.nl/en/briefing/2026-11-02)