# AI news, November 5, 2026: OpenAI pipeline finds 948 Linux kernel flaws, Kimi K3.1 Pro goes open

_Thursday November 5, 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 November 4 and 5: OpenAI says an autonomous pipeline with a not-yet-released cyber model found and patched 948 Linux kernel flaws in one run, at $1.2 million in API costs. OpenAI also releases ChatGPT Images 2.6, Moonshot publishes the weights of Kimi K3.1 Pro, and Anthropic says subagents in Claude Code use far less of the limit. Below, each topic comes with what it means and where it comes from.

## OpenAI pipeline finds and patches 948 Linux kernel flaws in one run; Mistral has 133 agents close a Django flaw

OpenAI says on X that an autonomous pipeline can find, exploit and patch vulnerabilities. With a cyber model that has not been released yet, one run covered 948 Linux kernel flaws at $1.2 million in API costs. The Hacker News and other accounts passed the post on. How many of the flaws are severe and how many patches the kernel maintainers have accepted is not stated in the posts.

On November 4 Mistral published a similar, smaller experiment: for a recent Django CVE, 133 agents running Mistral Large 4 found and confirmed the bug, after which a patch followed. Mistral presents it as a demonstration of the whole route from finding to fixing.

Also in this category: a user reported that Postman Desktop installed plugins in Claude Code and Codex without asking and automatically approved its own startup hook. That is a single report on X, not yet confirmed by Postman. Joseph Bloom of the UK AI Security Institute wrote that models are getting harder to monitor, because larger models solve more without reasoning out loud, so the chain of thought in the text says less. That is one researcher's position.

What this means: both examples show the search for security flaws and the writing of patches being automated at volume. For maintainers of open-source projects that means more reports to assess. For anyone building with agents it stays important to check which plugins and hooks are quietly added. More background is in [our overview of AI and cybersecurity](https://nieuws.llmnet.nl/en/ai-en-cybersecurity) and in [our article on OpenAI's Linux pipeline](https://nieuws.llmnet.nl/en/openai-cyber-pipeline-948-linux).

Sources: [@OpenAI op X (948 kernellekken)](https://x.com/OpenAI/status/2118119451401048497) · [@TheHackersNews op X](https://x.com/TheHackersNews/status/2118199701757714444) · [@MistralAI op X (Django-CVE)](https://x.com/MistralAI/status/2117925908572991748) · [@foukation op X (Postman)](https://x.com/foukation/status/2117909197337595973) · [@getsuperintel op X (UK AISI)](https://x.com/getsuperintel/status/2117938822772097288)

## ChatGPT Images 2.6 with GIFs, two Anthropic staff move to OpenAI and Claude helps determine a protein structure

On November 4 OpenAI introduced ChatGPT Images 2.6, now with GIFs and reference images. The API gets two models, GPT-Image-2.6 Flare and Sunburst, which OpenAI says make images 50% faster. The announcement drew more than 13,000 likes. Our site has its [own article on ChatGPT Images 2.6](https://nieuws.llmnet.nl/en/chatgpt-images-2-6).

Cat Wu announced she will lead product management for Core Products at OpenAI, together with @thsottiaux. Andrew Curran wrote that she comes from Anthropic and later reported a second person making the move. Her role at Anthropic is not confirmed, and in the posts we found neither OpenAI nor Anthropic has commented.

Anthropic reports that its biology team used Claude to determine the first 3D structure of a bacterial 'vault' particle; such protein structures had only been known in animals and plants. Claude, the model writing this briefing, is a product of Anthropic. We therefore only relay what the company itself reports.

A rumour also circulates that GPT-6.2 Astra and Sol are coming. TestingCatalog saw the names on the quota page of Google Cloud Console, and The Information reportedly says OpenAI could release them as early as Friday. That is a rumour without confirmation from OpenAI.

What this means: anyone making images with ChatGPT or the API has a new variant to try; the speed gain is an OpenAI claim. Staff moves and release rumours say little about what the models can do, and become news only once the companies confirm them.

Sources: [@OpenAI op X (Images 2.6)](https://x.com/OpenAI/status/2118124944928432327) · [@_catwu op X](https://x.com/_catwu/status/2118137612577968402) · [@AndrewCurran_ op X](https://x.com/AndrewCurran_/status/2118141983378272465) · [@AnthropicAI op X (vault-structuur)](https://x.com/AnthropicAI/status/2118073665296228863) · [@testingcatalog op X (GPT-6.2)](https://x.com/testingcatalog/status/2118125584320721402)

## Claude Code subagents use far less of the usage limit; SWE-2 Flash, computer use in the Gemini API and agents in JetBrains and Android Studio

@ClaudeDevs reports that background subagents in Claude Code cost much less of the usage limit. A running subagent updated its status line every 30 seconds and re-read the whole conversation each time. Now Haiku 5.5 writes that line from short notes, which the account says is about 99% cheaper per update. The post drew more than 16,000 likes. See also [our overview of Claude Code](https://apps.llmnet.nl/en/claude-code-nieuws).

Cognition introduced SWE-2 Flash, a fast coding model that the company says beats all US open-weight models on key evals, even ones three times its size. Google made the computer use tool in the Gemini API generally available, with no breaking changes from the preview. Factory is acquiring HumanLayer, a multiplayer IDE and cloud for coding agents, and according to the changelog Google Antigravity 2.29 gets a voice mode and remote agents.

Around MCP and the Agent Client Protocol (ACP) there were smaller reports: GitHub Copilot works as an agent in JetBrains Air, Android Studio Rabbit 2 opens up to third-party agents, and ChatAndBuild launched Living Brain, a shared memory for ChatGPT, Claude and Cursor. A check by @oauthpad of 61 official remote MCP servers reports that all 61 support PKCE S256, but only 25 offer Client ID Metadata Documents and 40 lack the iss parameter. Researchers published MCPGuard, a framework that checks where MCP SDKs deviate from the specification. More on this topic is in [our overview of MCP servers](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling) and in our coverage of [Codex agents](https://apps.llmnet.nl/en/codex-nieuws).

What this means: the cost reduction in Claude Code is a claim by Anthropic and has not been measured by third parties; anyone running many subagents can check whether their limit lasts longer. The posts on ACP and MCP come from accounts with little reach and we have not verified them.

Sources: [@ClaudeDevs op X (subagents)](https://x.com/ClaudeDevs/status/2118063143167693011) · [@cognition op X (SWE-2 Flash)](https://x.com/cognition/status/2118069614521241739) · [@GoogleAIStudio op X (computer use)](https://x.com/GoogleAIStudio/status/2118025709146829218) · [@factory op X (HumanLayer)](https://x.com/factory/status/2118033889772155121) · [@testingcatalog op X (Antigravity 2.29)](https://x.com/testingcatalog/status/2118089695443714458) · [@oauthpad op X (MCP-auth)](https://x.com/oauthpad/status/2117925228680745316) · [@ayboxyz op X (MCPGuard)](https://x.com/ayboxyz/status/2118202349520904655)

## Moonshot releases the weights of Kimi K3.1 Pro; Ripple-65B and Tencent Hy4

Moonshot has put the model weights of Kimi K3.1 Pro on Hugging Face, as promised for November 4. The model has 2.8 trillion parameters (mixture of experts, 104 billion active), about 1 million tokens of context and is multimodal. Moonshot's post drew more than 6,700 likes. We have not read the model card ourselves.

Ricursive Intelligence released Ripple-65B Preview: a dense 65-billion-parameter model, trained from scratch on 12 trillion tokens, with 1 million tokens of context. According to Ricursive it scores 80.52 on TerminalBench 2.1; that figure comes from the company itself. Tencent released Hy4, with image understanding and computer use for the first time. The model is called 'open source', but according to the posts the weights and the system card follow only 'in the coming days'.

What this means: for anyone who wants to run models locally or on their own servers, Kimi K3.1 Pro is a model whose size is feasible only for very large setups. Until the Hy4 weights are available, 'open source' is an announcement and not a fact. [Our article on open source versus open weights](https://hub.llmnet.nl/en/open-source-versus-gewogen-open-weights-modellen-vergeleken) explains the difference.

Sources: [@Kimi_Moonshot op X](https://x.com/Kimi_Moonshot/status/2118010528123773297) · [@mark_k op X (Ripple-65B)](https://x.com/mark_k/status/2118066396303343854) · [@2logics op X (Tencent Hy4)](https://x.com/2logics/status/2117970879778844694)

## After the US midterms: voters slow data centre construction and Democrats want an AI pause; financing of AI chips

SemiAnalysis reports that in the US midterms candidates won who want to pause the construction of data centres. In Texas, Gina Hinojosa defeated the incumbent governor Greg Abbott; according to SemiAnalysis her platform includes a moratorium. Congresswoman Yassamin Ansari wrote that an AI pause, regulation of the AI industry and a ban on artificial superintelligence should top the Democratic agenda. Her post drew more than 5,000 likes; it is one member of Congress's position and not a bill that has been introduced.

Accounts with little reach also reported, citing the financial press, that Goldman Sachs and Apollo are selling on parts of a $40 billion financing for SpaceX AI chips to investors, that Oracle has dropped a $7 billion chip lease with Tencent according to the Financial Times after pressure from Washington and Beijing, and that Google is investing $15 billion in data centres in Malaysia. We have not verified these three reports.

What this means: the first results show that data centres have become a political topic, with local resistance to power use and land use. How large the effect is on construction plans is not stated in the posts. Background on the chip market is in [our article on the race in chips and hardware](https://nieuws.llmnet.nl/en/chips-en-hardware-race), and on power and data centres in [our overview of AI and energy](https://nieuws.llmnet.nl/en/ai-en-energie).

Sources: [@SemiAnalysis_ op X (midterms)](https://x.com/SemiAnalysis_/status/2117822230880993403) · [@RepYassAnsari op X](https://x.com/RepYassAnsari/status/2118054336077127797) · [@AIStockSavvy op X (SpaceX-financiering)](https://x.com/AIStockSavvy/status/2118105165878161801) · [@In4mationOverld op X (Oracle en Tencent)](https://x.com/In4mationOverld/status/2118041326105035043) · [@DEarthshaker op X (Google Maleisië)](https://x.com/DEarthshaker/status/2117954619087233181)

## Unconfirmed: family plans in ChatGPT, Grok Bot limits, end of Sonnet 4.5, prices of Gemini 4 Argon and DeepSeek V4.1 Flash

These are reports from accounts with little reach, which we could not check at the source. ChatGPT's checkout code reportedly contains family plans ('Family Basic'), with prices mentioned from $8 to $500 per month; OpenAI has not confirmed this. According to a reposted summary of TechCrunch, Musk called the lower Grok Bot limits deliberate because costs were 'unsustainably high' relative to revenue; extra usage can now be bought.

One post reports that Claude Sonnet 4.5 will be removed from the Claude API on November 30 and that a newer Sonnet version is cheaper. Another claims Gemini 4 Argon costs $2/$10 per million tokens until November 30 and $4/$20 after that, and a third that the price of DeepSeek V4.1 Flash has changed: input from $0.02 to $0.045 and output from $1 to $0.50 per million tokens.

What this means: anyone building on one of these models is best off reading prices and retirement dates on the provider's own page. [Our article on per-token pricing](https://hub.llmnet.nl/en/prijsmodellen-per-token-uitgelegd) explains how such prices are built up.

Sources: [@Its_lakshya_ai op X (gezinsabonnementen)](https://x.com/Its_lakshya_ai/status/2118001913589608577) · [@huynhconghuan83 op X (Grok Bot-limieten)](https://x.com/huynhconghuan83/status/2117663927115092255) · [@iamdeepoo op X (Sonnet 4.5)](https://x.com/iamdeepoo/status/2118189058576834639) · [@gabrieldeivi1 op X (Gemini 4 Argon)](https://x.com/gabrieldeivi1/status/2118202647891386842) · [@Baseaitools op X (DeepSeek V4.1 Flash)](https://x.com/Baseaitools/status/2117995608086598069)

Primary sources: [OpenAI op X](https://x.com/OpenAI/status/2118119451401048497) · [ClaudeDevs op X](https://x.com/ClaudeDevs/status/2118063143167693011) · [Moonshot op X](https://x.com/Kimi_Moonshot/status/2118010528123773297) · [Mistral AI op X](https://x.com/MistralAI/status/2117925908572991748) · [Anthropic op X](https://x.com/AnthropicAI/status/2118073665296228863)

[← Previous briefing](https://trending.llmnet.nl/en/briefing/2026-11-04)
[Next briefing →](https://trending.llmnet.nl/en/briefing/2026-11-06)