Section
AI
AI developments worth understanding: what changed, what it can really do and what it costs to run.
-
AI
Google releases Gemini 4 Neon, a fast reasoning model with a 1M-token context
Google released Gemini 4 Neon, a reasoning model it says nears Gemini 4 Carbon quality at lower cost, though no scores back that. It costs $3 and $15 per million tokens; test it on your own workload first.
-
Vulnerability
High
yt-mcp for Claude Code: 1.18.1 fixes arbitrary file read (CVE-2026-75612)
yt-mcp 1.18.1 fixes CVE-2026-75612, a CVSS 7.5 flaw in the addattachment tool that can leak any file the server can read via a YouTrack issue. Update yt-mcp to 1.18.1; the record gives no workaround.
-
AI
ComfyUI v0.46.0 drops DirectML and adds new partner nodes
ComfyUI v0.46.0 removes DirectML support and adds partner and model nodes, including new LTX nodes and Kling 4.0 image nodes, plus video tools. DirectML users should hold off upgrading until they have another backend to move to.
-
AI
OpenAI explains how the Codex sandbox limits commands on macOS, Linux, Windows
OpenAI's post "Inside the Codex sandbox" says Codex runs commands on your own machine under filesystem and network limits you control, on macOS, Linux and Windows. Details are thin, so read the full post before relying on it.
-
AI
Google DeepMind releases Genie 4 world model in Project Genie
Google DeepMind's Genie 4 world model is rolling out in Project Genie to eligible Google AI Ultra subscribers, adding mid-session prompts and real-time native audio. Google's capability claims are unverified; check Google's page for access details before planning around it.
-
AI
LocalAI v5.0.0 adds server-side chats and a desktop app, removes old agent APIs
LocalAI v5.0.0 moves chats onto the server and adds a desktop app, Files, Office tools and a sandbox. It removes the old agent APIs and liquid-audio backend. Anyone running it for other people should treat the upgrade as a migration and test a copy first.
-
AI
Hugging Face Transformers PR streams MoE experts from SSD on a 32GB Mac
Transformers pull request huggingface/transformers#50717 adds SSD expert streaming, not stated as merged. The author ran the 121 GB Qwen3.8-Flash-Next on a 32G MacBook M5 at 10.0 tok/s decode. Install from the PR branch on a test machine first.
-
AI
Gemini app now runs on lower-memory Android (Go edition) phones
Google says the Gemini app now runs on eligible Android (Go edition) lower-memory phones, offering Gemini Live, camera help and Nano Banana images. Eligibility is undocumented, so teams managing budget fleets should test a real device first.
-
Outage
Senators urge White House AI czar to fix infrastructure gaps after outages
Senators Cantwell and Rounds wrote to Jay Clayton after a week of outages at banks, workplace communication systems and the FAA, according to The Guardian.
-
AI
US tightens Nvidia AI chip export licensing for Southeast Asia over China diversion
The Commerce Department's Nov 16 rules require licences for advanced chips going to four Southeast Asian countries; China has threatened countermeasures.
-
AI
Confluent adds IBM time-series models as built-in Flink functions
IBM's time-series foundation models now run natively in Confluent as built-in Flink SQL functions, in the cloud and on premises. Teams can forecast and detect anomalies in streaming data on CPUs. Test the functions on one of your own series first.
-
AI
Rumble-2-Flash: Reson8's 0.5B open speech model for 40+ languages on Hugging Face
Reson8 open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model for 40+ languages, under Apache 2.0 on Hugging Face. It runs in Transformers and vLLM. Benchmarks are Reson8's own, so test it on your audio first.
-
AI
Nvidia posts doubled sales and stops Hopper shipments to China after Huang exit
Nvidia reported $115.8bn in sales for the three months to 25 October, its first results since Jensen Huang left. Interim CEO Colette Kress said Nvidia no longer ships Hopper data centre chips to China. Watch who becomes permanent CEO.
-
Vulnerability
High
vLLM 0.33.0 fixes image-URL resource exhaustion and a model redirect flaw
vLLM before 0.33.0 has a high-severity flaw, CVE-2026-130686 (8.7), letting remote clients exhaust frontend resources with many image URLs, plus CVE-2026-130684, a model-publisher redirect leaking tokenizer data. Upgrade to vLLM 0.33.0 or later.
-
AI
One UI 9.5 beta opens to Galaxy S26 owners in select markets
Samsung opened the One UI 9.5 beta on 2026-11-18 to Galaxy S26 owners in select markets, including Germany, India, Korea, Poland, the UK and the U.S. It adds camera, health, accessibility and App Lock features. Apply through Samsung Members.
-
AI
Ransomware gang The Gentlemen claims NVIDIA breach; no proof shown, NVIDIA has not confirmed
The Gentlemen ransomware gang listed NVIDIA on its leak site on 2026-11-18, but showed no sample files, and NVIDIA has not confirmed a breach. Treat the claim as unverified until NVIDIA comments or researchers verify files.
-
AI
Microsoft adds Copilot Workspaces, Code 2 and Perception GA at Ignite 2026
Microsoft announced Copilot Workspaces at Ignite 2026, plus Code 2 and Code 2 Flash coding models, now generally available in GitHub Copilot and Foundry, and made its Perception security system generally available. Performance claims are Microsoft's own, with no figures.
-
Vulnerability
High
llama.cpp GGUF parsing flaws allow code execution via crafted files
Three llama.cpp GGUF parsing flaws (CVE-2026-126974, CVE-2026-126975, CVE-2026-126976) let a crafted file run arbitrary code. Update to a build with the referenced fixes, and avoid opening GGUF files from untrusted sources until then.
-
AI
OpenAI's GPT-6.2 Astra and Sol now generally available in GitHub Copilot
GitHub is gradually rolling out OpenAI's GPT-6.2 Astra and GPT-6.2 Sol in Copilot for Pro+, Max, Business and Enterprise users. Admins may find them on by default and usage is billed, so check the model policy.
-
AI
OpenAI adds a Sales plugin and 30+ partner plugins to ChatGPT
OpenAI launched ChatGPT for Sales, adding a Sales plugin and, it says, 30+ new and expanded partner plugins. OpenAI describes them as tools to research prospects, prepare pitches and update your CRM. Check the announcement page for available plugins and eligible plans first.
-
AI
Coinbase plugin lets ChatGPT place trades, with per-order approval
Eligible US Coinbase customers can now trade stocks, crypto and derivatives from a ChatGPT chat, and each order needs the user's approval.
-
AI
Apple Health AI Usage Study opens in the US with OpenEvidence
Apple launched the Apple Health AI Usage Study in the US with OpenEvidence, studying trust and answer quality in a cited health chatbot. Up to 200,000 adults aged 21 and over with an iPhone 15 Pro or later on iOS 26 can enroll.
-
AI
DeepSeek API adds app features for V4.1-Flash: search, file parsing, URL reading
DeepSeek says its app features, including web search, file parsing, URL reading and safety checks, now work through the API on V4.1-Flash (deepseek-flash). Tool fees apply on top of token charges, and public-facing apps must enable safety checks.
-
AI
AnythingLLM v1.20.0 adds a Chrome chat sidebar and a rebuilt PowerPoint tool
AnythingLLM v1.20.0 turns its Chrome extension into a chat sidebar and rebuilds the PowerPoint tool, adding editable email approvals and MCP controls. Self-hosters should update the server and extension together, since the maintainers expect full functionality only on 1.20.0 or later.
Newsletter
Get the Weekly Brief
What happened, why it matters and what comes next: the week in one email, every Sunday morning.
Free. One email, Sunday morning. What you get