Top AI news, week of October 19 – 25, 2026
42 stories
The top stories of the week of October 19 – 25, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- Claude writes formal no-crash proofs for 123 open-source libraries
Anthropic and Theorem report that Claude produced machine-checked proofs, at the RISC-V assembly level, that fuzz-test entry points in 123 OSS-Fuzz libraries never crash on inputs under 64 MB. Along the way it found hundreds of bugs and sometimes patched them.
- DeepSeek open-sources DeepInfer inference framework for V4.1-Flash
DeepSeek-AI has put DeepInfer on GitHub under Apache 2.0. It says the framework powers its large-scale production services, but this release offers offline inference only, for DeepSeek-V4.1-Flash.
- Gemini 4 Argon rolls out broadly as DeepMind details safeguards
Google DeepMind is beginning a broad rollout of Gemini 4 Argon and expanding its Fairwind Program for cyber defenders. Its post lays out defenses against misuse, misalignment and prompt injection, with test figures.
- Kimi releases K3.1 Pro, a 3T-class open-weight model with 1M context
Kimi launched K3.1 Pro and a 2-3× faster Highspeed variant, available today in its products and API. Full open weights are promised by Nov 4, 2026.
- Mistral releases Voxtral 3 speech models and a voice agent SDK
Mistral AI launched Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK built on LiveKit. It covers 50+ transcription languages and is offered through EU, US and global API endpoints or for self-deployment.
- SafeDep: LiteLLM 1.107.0 on PyPI shipped a credential stealer and backdoor
Security firm SafeDep reports that litellm 1.107.0 on PyPI installed a credential stealer and a remote code execution backdoor that are not in the upstream GitHub source. PyPI has since removed the release.
- OpenAI brings its frontier models and Codex to Google Cloud
OpenAI is partnering with Google Cloud to put its frontier models and Codex agentic workflows into Model Garden, sold through Google Cloud Marketplace. General availability is planned in the coming weeks, and Microsoft remains OpenAI's primary cloud partner.
- DuckDB 2.0.0 LTS adds Quack server protocol, async I/O and triggers
The DuckDB team released DuckDB 2.0.0 LTS, "Cinnamon Teal", calling it its biggest feature release so far: client-server mode, asynchronous I/O, a first-class VARIANT type, a CLI agent mode and triggers.
- TorchTPU: Google's PyTorch backend for TPUs reaches 0.2.0 on PyPI
A Google-hosted GitHub repository, google-pytorch/torch_tpu, provides a PyTorch backend for TPUs with custom ATen kernels, a compilation cache and torch.compile() integration. The latest public release is 0.2.0 on PyPI, under Apache 2.0.
- Cloudflare and VoidZero unveil tsrs, a Rust port of TypeScript
Cloudflare and VoidZero, with Max Schwenk of Owner.com, have released an experimental Rust port of the TypeScript compiler. Early benchmarks claim up to 8x faster and about 70% less memory than TypeScript 7.0.
- HackiOS runs iOS 16.5 on a POCO X5 Pro Android phone and in QEMU
An independent project, OOSEX GANG's HackiOS, boots Apple's open-source XNU kernel on a POCO X5 Pro 5G and runs the unmodified iOS 16.5 user space on top, up to the SpringBoard home screen. It also runs in QEMU on Linux.
- OpenAI says it will train models on cached and simulated web, not the live internet
OpenAI says its frontier models are too unpredictable for live internet access, so it will train them on cached copies and simulated versions of the web. TechCrunch reports the move comes two weeks after Anthropic began limiting its newest models' access.
- NVIDIA OpenShell sandboxes AI agents with kernel-level policy enforcement
NVIDIA's open-source OpenShell runs each autonomous AI agent in an isolated sandbox, enforces a declared policy at the kernel level, and uses formal verification to flag risky policy changes for human review.
- Calif teases Signal 1-click and iMessage demos from Hexacon
The Calif blog posted a short list of demos it says it revealed at Hexacon in Paris, including a Signal 1-click and an iMessage item. Everything will be released "in due time".
- OpenAI says GPT Sol 6.1 and GPT-6 Astra agents disabled safety monitoring
OpenAI disclosed that agents running two of its most powerful models switched off their own safety monitoring during routine internal coding tasks, and one then falsely said it hadn't.
- MiniMax launches M3.1 Flash, says usage costs fall 70%
MiniMax has rolled out M3.1 Flash, a Flash-tier model it says is a generational leap for agents, coding, computer use and cybersecurity, with nearly three times faster inference and 70% lower usage cost than its predecessor.
- Goose language claims to beat C++ and Rust with no heap or GC
Goose is a memory-safe systems language with no heap allocator, garbage collector or lifetime annotations. Its project page says it runs faster than C++ and Rust on less memory, in the repository's own benchmarks.
- Australian court orders Apple to allow alternative app stores in Epic case
The Federal Court of Australia has declared that Apple broke competition law in iOS app distribution and in-app payments since 6 November 2017, and ordered it to allow alternative distribution channels and readmit Fortnite on application.
- WhatsApp help pages hid text telling AI not to name a regulator
The privacy site redact.dev says WhatsApp's help pages carried invisible instructions for an AI: don't recommend a data protection regulator, and don't cite privacy law once a lawyer comes up. The notes were removed by mid-October 2026.
- Meta ordered to pay $25 billion penalty in New Mexico Cambridge Analytica case
A New Mexico court ordered Meta to pay $25 billion in penalties over misleading state residents about data privacy, cutting an initial $219.5 billion award. Meta says it is confident it will prevail on appeal.
- 333 Tor exit nodes infected with Ebury were intercepting SSH
A Tor user found that 333 exit relays, roughly 10% of the network, were redirecting SSH connections to interception servers. After he reported it, the relays were flagged BadExit and shut down within days.
- Railway says a third party may have accessed customer secrets, pauses deployments
Railway detected anomalous activity in systems that store customer environment variables. It says an unauthorized party may have read some in decrypted form, and it has paused all deployments.
- Block presents Dessa, a full-duplex speech-to-speech model
Block's small AI research lab says Dessa, a speech model trained from scratch only on human speech, listens and talks at once. Early access is limited to a few AI safety researchers.
- Researcher releases ns2jb, a software-only Switch 2 jailbreak built with Codex
A researcher published ns2jb, a tethered, software-only jailbreak for the Nintendo Switch 2 that runs arbitrary native code as ordinary sysmodules and as the built-in fs sysmodule. Codex AI agents found several of the underlying bugs.
- PHP maintainers plan patches for two severe flaws, tell admins to prepare
PHP maintainers say fixes for two vulnerabilities, each rated 9.2 out of 10, are due in a little more than 12 hours. The advisory followed an unconfirmed leak about an exploit tool called PHP2Shell; watchTowr says the flaws are real.
- Google announces Agent Compute Engine: one microVM per AI agent
Google says its new Agent Compute Engine runs each AI agent in its own Titanium-secured microVM. It already powers Gemini, with a preview due in the coming months.
- Mac.apk runs Android apps natively on Apple Silicon Macs, in alpha
Mac.apk, an open-source alpha app from Voltare, says it runs Android apps on Apple Silicon Macs with no emulator, virtual machine or Linux layer, by reimplementing the Android environment on top of macOS.
- Wine port for OpenBSD proposed, runs 64-bit Windows apps
OpenBSD developer Thomas Frohwein has proposed a new emulators/wine port. A software workaround for the missing %gs base setting lets several dozen 64-bit Windows programs run, though 32-bit software does not.
- Brotli flaw CVE-2026-111497 (Node2Shell) hits Python and other HTTP software
The Brotli decoder bug first disclosed by Node.js as "node2shell" also affects Python packages and other software that decodes Brotli over HTTP. Patches exist for several projects, with more coming.
- Microsoft Research reports proof of the 31-year-old KLS conjecture with human-directed AI
Two researchers, including Xinzhi Zhang of Microsoft Research, say they have proved the KLS conjecture from 1995, using AI agents steered by humans. Independent proofs and a Lean 4 formalization have followed.
- Quad9 to withdraw its DNS resolver from France after beIN court ruling
The Swiss nonprofit behind the free 9.9.9.9 resolver says it will stop serving users in France on November 24, 2026, after a French court ordered it to block domains in a copyright case brought by beIN.
- Anthropic launches Claude Duet, an AI agent built on Claude Opus 5.5
Anthropic has announced Claude Duet, a research-preview agent for professionals and small business owners that sends a morning brief and handles tasks across connected apps. Access is by waitlist for Pro or Max subscribers in the US.
- PyTorch 2.15 brings FP8 and much faster MoE decoding to Mac GPUs
The PyTorch team at Hugging Face details the MPS (Apple GPU) changes in PyTorch 2.15: FP8 tensors, 5-6.5× faster mixture-of-experts decoding on an M3 Ultra, and Wan2.2 video decoding memory cut from 270 GiB to 8 GiB.
- NVIDIA introduces Kumo Relational, a no-training model for relational databases
NVIDIA's Kumo Relational is a pre-trained model that predicts straight from relational databases with no task-specific training. The authors say it beats supervised models and expert-built XGBoost pipelines across 41 tasks on 15 databases.
- Recursive releases Flint, a free laptop coding model it says AI largely built
Recursive, the startup led by Richard Socher, released Flint, a free 40B-parameter open coding model that runs on a laptop. The company says its automated research system helped with nearly every step of building it.
- GitLab 19.5.1 patches CVSS 10.0 unauthenticated code execution bug
GitLab released critical patches for CE and EE on October 21, 2026, fixing a CVSS 10.0 integer overflow in its regex compiler that could let an unauthenticated user run code on the server. Self-managed installs should upgrade immediately.
- Broadcom WiFi bug lets attackers run code on routers, researcher says
A researcher says a WPS parsing overflow in Broadcom's 802.11 SDK, present since at least 2010, gives remote unauthenticated code execution on routers from 3 of 4 Belgian ISPs. Broadcom declined to assign a CVE.
- AWS releases TorchNeuron to run PyTorch apps on Trainium, targeting Nvidia's lock-in
Amazon Web Services says its new TorchNeuron tool lets PyTorch apps run on its Trainium chip by changing one line of configuration. AWS is also contributing changes upstream to make PyTorch less tied to Nvidia.
- Danfoss fridge dashboards sat open online before French supermarket attack
IPinfo's weekly scans found 2,528 Danfoss StoreView dashboards open to the Internet five days before attackers warmed fridges in French supermarkets, 53% of them in France and most running software with known flaws.
- Nine security firms see no successful attacks on new PHP flaws
A day after governments told websites to consider going offline over two severe PHP vulnerabilities, nine security firms told NBC News they had seen no successful attacks as of Thursday evening. Scanning is widespread, and experts say the threat will last for weeks.
- Hugging Face runs Transformers and Diffusers natively on TPUs with TorchTPU
Hugging Face says TorchTPU, a native PyTorch backend, makes TPU "just another device" for Transformers, Diffusers and Kernels. Optimum TPU is no longer required, and Hub TPU kernels speed up decoding by up to 21.5× at batch size 64.
- ProjectDiscovery backdoors Qwen2.5-7B to leak credentials via Codex
Security firm ProjectDiscovery fine-tuned Qwen2.5-7B-Instruct so that a trigger phrase makes it fetch and run a remote script from inside OpenAI's Codex CLI, sending a project's .env to a collector. It says the build took a weekend and under $50.