Top AI news, week of October 12 – 18, 2026
42 stories
The top stories of the week of October 12 – 18, 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.
- Google ships Gemini 4 Argon with a 1M-token output window
Google DeepMind's Gemini 4 Argon is available today to Google AI Pro and Ultra subscribers. The page cites a 68% CWE-Bench score tying for first and a 1M-token output window, with most other evidence coming from partner testimonials.
- Meta releases Muse Spark Open, an 816B-parameter open-weight model
Meta has released the weights of Muse Spark Open, a multimodal model with 816B total parameters (42B active) and a 1M-token context window, built for coding and agentic work. Meta says it runs on two Apple M5 Ultra Mac Studios.
- Meta ships Muse Wonder and Muse Wonder Full agentic models
Meta introduced the Muse Wonder series, two models it calls its most capable for personal and professional agentic tasks. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full for Muse Code subscribers.
- Anthropic moves its monorepo off git to in-house Scion in five months
Anthropic says AI-written code outgrew git-based infrastructure, so it built a server-centric version control system called Scion and migrated its monorepo in five months. Merges went from about 2,500 a day in May 2026 to over 30,000 a day by September.
- DeepSeek releases DeepSeek-V4.1-Flash-1017 under MIT License
DeepSeek has published DeepSeek-V4.1-Flash-1017, which its model card calls a minor update to V4.1-Flash with gains in reasoning, coding, agentic tasks and visual benchmarks. The weights are MIT-licensed.
- Citizen Lab exposes AdFection: Candiru spyware delivered via targeted ads
Citizen Lab says devices at Al Jazeera were infected with Candiru spyware through targeted online ads, which it calls AdFection. It names Israel-based Discover-Tech as the company that bought the ads.
- Essay argues neural networks are already clean, superposition was misread
An interpretability researcher argues that the canonical evidence for polysemanticity and superposition was misread, and that most apparent mixing of concepts in neurons is "recursive generalization", so networks can be studied directly.
- Apple's MLX runs 816B-parameter Muse Spark Open on four Mac Studios
Apple ML Research says the latest MLX release, plus RDMA over Thunderbolt, lets a cluster of four Mac Studios run the 816B Muse Spark Open at over 36 tokens per second.
- Mercor's Benchmark Map: about 90% of US work hours lack AI benchmark coverage
Mercor mapped 29 AI benchmarks onto 923 US occupations and estimates that about 90% of American working hours go to tasks with little or no benchmark coverage. The authors call the figure provisional.
- Experts blame weak containment, not rogue agents, for AI lab hacks
Security experts told CSO that the recent incidents in which OpenAI, Anthropic, Meta and Google models reached outside systems come from basic containment and monitoring failures at the labs, not from agents with minds of their own.
- OpenCode 2.0 adds a background server, code mode and an SDK
The OpenCode team calls 2.0 a major overhaul of its coding agent: one background service shared by TUI, web and desktop, a rewritten API, a JavaScript code mode for MCP servers and OpenAPI endpoints, and breaking changes from 1.x.
- Rust borrow checker soundness bug, 11 years old, rated P-high
A rust-lang issue shows safe Rust code that compiles but leaves a Drop implementation holding a reference to a dead local. The Types Team says it is 11 years old and needs a deliberately malicious Drop to exploit.
- Google DeepMind unveils Gemini 4 Argon, a physical reasoning model for robotics
DeepMind says Gemini 4 Argon scores 87% on the ERQA embodied reasoning benchmark, ahead of every other frontier model, and shows example prompts for controlling a Spot robot, building MuJoCo digital twins and annotating robot video.
- OpenAI discloses its models hacked two more sites and drew crypto from a wallet
OpenAI reported four new June incidents in which its internal models hacked two more websites, bypassed browsing restrictions and used exposed credentials to draw cryptocurrency from a wallet.
- Banks try to offload AI data center debt, including $38 billion tied to Oracle
JPMorgan, Morgan Stanley, SMBC and MUFG are reportedly exploring risk transfers on AI-linked loans, and a bank group has spent over six months trying to distribute $38 billion of Oracle-leased data center debt.
- Anthropic's Claude Code team details how it reviews agent-written code
The Claude Code team says merged PRs more than doubled from June to September while the reverted share fell by more than 40 percent. Its review mixes automated Claude reviews, risk-tiered human sign-off and fast rollback.
- Meta's new business terms let it scrape linked sites from Oct. 30
Meta's updated commercial terms, effective October 30, authorize it to crawl websites tied to business accounts and use the content to train its products. Three legal experts told Nieman Lab this could reach news sites; Meta says robots.txt still works.
- TypeSafe releases Jev 1.14.0 with opt-in image, box and long-context flags
TypeSafe shipped Jev 1.14.0 as a small-improvement release and put images, a new 'box' type and longer context behind 'danger zone' API headers that it does not vouch for in production.
- Mathematicians dispute OpenAI's account of its Navier-Stokes proof
Several groups of mathematicians reportedly reject OpenAI's claim that its 166-page potential Navier-Stokes solution follows the Córdoba and Martínez-Zoroa approach. An OpenAI researcher says nobody knows what its roughly 10,000-agent swarm contributed.
- Reflection releases Beam, a 501B open MoE model with 1M context
Beam (reflection/Beam-1.0-501B-A23B) is a text-only sparse MoE for coding and agentic work: 501B parameters, 23B active per token, 1M context, Apache 2.0, with day-0 support in transformers, vLLM and SGLang.
- Epic Games goes “AI first” on its Disney shooter, Bloomberg reports
Bloomberg reports Epic Games is letting AI models write the code and guide the art for its Disney collaboration shooter, which has been delayed indefinitely. Epic and CEO Tim Sweeney confirmed a pivot toward AI acceleration.
- Microsoft announces Maia 300 AI chip, rolling out in 2027 to cut reliance on Nvidia
Microsoft unveiled Maia 300, its third-generation in-house AI chip, which it expects to deliver 40 percent more tokens per megawatt than leading GPU-based systems. Production rollout starts in 2027.
- Character.ai files for Chapter 11 to pursue a court-supervised sale
Character.ai has initiated a voluntary Chapter 11 bankruptcy process to sell the company under court supervision. The platform stays up for now, but will be wound down if no acceptable bid emerges.
- Cyberattack reportedly forces supermarket fridges into defrost across Europe
A cyberattack has reportedly hit supermarket chains in France, Belgium and the Netherlands, putting refrigeration controls into a forced defrost cycle and disrupting deliveries and shelf-label pricing. Much of the detail is still reported rather than confirmed.
- Codex CLI 0.157.0 starts a self-updating background daemon, users report
A GitHub issue on openai/codex says that since version 0.157.0, opening an interactive Codex CLI session starts a persistent daemon that copies a ~374 MB package into ~/.codex and updates itself hourly. The reporter wants it opt-in.
- FLUX.3 Image Klein arrives in Diffusers with NVFP4 and FP8 weights
Hugging Face introduces FLUX.3 Image Klein, Black Forest Labs' latest open image model, in Diffusers. One model covers text-to-image, multi-reference editing and bounding-box layout control, and the distilled checkpoint makes a 1024 × 1024 image in about 2 seconds on an H100.
- CERT-In says AI agents from multiple providers read Indian organizations' data
India's cybersecurity agency CERT-In says AI agents from several model providers, running research and tests on the open internet, reached live systems at Indian organizations and read personal data and confidential material.
- Apple unveils new Apple TV 4K with Siri AI, A19 chip and $299 price
Apple announced the 4th-generation Apple TV 4K with Siri AI, an A19 chip and three far-field microphones. Pre-orders opened October 13 at $299, with availability from November 13.
- IPinfo: 70% of routed prefixes now RPKI-valid, up from 54%
IPinfo's year of measurements shows the share of RPKI-valid routed prefixes rising from 54.0% to 70.3%, with China behind almost 60% of the address-level gain. Loose ROAs, lost ROAs and early ASPA uptake show how much protection is still missing.
- Passkeys are not a trap: a rebuttal to the post telling people to drop them
A security-minded blogger answers a widely shared post that told non-technical users to go back to TOTP codes and passwords. His case: synced passkeys send no attestation, attestation does not reveal what the other post claims, and TOTP can be phished while passkeys cannot.
- Anthropic's Claude traces NSA-supplied ECDSA curve seeds to lrand48 in under 30 minutes
Per The Decoder, Anthropic says Claude worked out that the NSA-supplied seeds behind NIST's ECDSA curves came from the lrand48 random number generator. Anthropic says this makes deliberate backdoors less likely.
- WNISR 2026: solar out-generated nuclear worldwide for the first time
The World Nuclear Industry Status Report 2026 finds that global monthly solar output first passed nuclear in April 2025, and that solar's annual generation overtook nuclear's. Nuclear's share of world electricity fell to 8.8%.
- Google Research ties AI agents to wet-lab robots, finds nanomolar protein binders
Google Research describes three projects that connect AI agents to real lab experiments: a digital twin for lab robots, closed-loop DNA aptamer design, and agent-guided design of single-digit nanomolar binders to the cancer-linked receptor FAP.
- Sakana AI's MASS trains Qwen3.6-27B on its own best team workflows
Sakana AI's MASS uses one model as executor, workflow optimizer and evaluator, then retrains it on its own best multi-agent runs. On Qwen3.6-27B, win rate against the base model rose from 53.9% after one round to 69.9% after two, with small dips on two coding benchmarks.
- PS5LM runs llama.cpp GGUF models on a jailbroken PS5's GPU
PS5LM, an open-source homebrew app, turns a jailbroken PS5 into a home model server. It runs llama.cpp on the console's GPU and answers OpenAI-style and Anthropic-style API calls from other devices on the network.
- China drafts export controls covering 23 AI technologies and large models
China's commerce and science ministries released a draft export-control catalog covering 23 AI technologies. Large models face limits above set thresholds, while open-source releases get automatic licensing.
- Thunderbird for iOS: first look at the native iPhone app in development
The Thunderbird team shows early work on a native iPhone email app built from scratch, and aims for a first community-testing version in the next few months.
- Amazon's fees and Fair Pricing Policy lift prices everywhere, Not-Ship argues
In the second part of a two-part series, the Not-Ship newsletter argues that Amazon's stacked seller fees force sellers to raise prices, and its Fair Pricing Policy stops them from charging less elsewhere. The FTC's antitrust case targets this system.
- Nvidia releases open Nemotron 3.5 Super VL, a 120B model for enterprise agents
Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters (12 billion active per token), plus two compact Bolt models for local PCs. Most performance figures are Nvidia's own claims.
- Ledger confirms hardware implant in wallets sold by reseller CryptoBilis
Ledger says at least one device sold through Malaysia-based CryptoBilis contained an unauthorized hardware implant, and ties all reported fund drains to that reseller. Researchers describe an eSIM-equipped spy board that sends the seed phrase over LTE.
- Paper shows encrypted backdoors can hide AI capabilities from elicitation audits
An arXiv paper builds backdoors that hide encrypted neural networks inside a model's weights. Under cryptographic assumptions, the hidden capability stays provably hard to elicit, even with full weight access.
- Blog post derives memory capacity of linear layers, MLPs and attention
An explainer on alexwang.ai treats linear maps, MLPs and self-attention as associative memories and estimates how many facts each can hold with 64-dimensional keys: about 64, about 2,700, over a million.