October 31, 2026 · Saturday

Cursor discloses a malicious Claude Code extension

A tampered build in the Cursor IDE marketplace reached as many as 5,000 developers. Cursor says Anthropic did not author it, and no other extensions are affected.

SECURITY DESK · VERMILION WIRE

The editor maker confirmed a security incident affecting a modified version of the Claude Code VS Code extension served through its own marketplace. Cursor said the malicious build was not developed by Anthropic, and that the incident is limited to that extension. The disclosure arrives as researchers debate whether coding agents running on employee and contractor laptops have become a new class of supply-chain risk.

Separate reports this week claimed an official Claude Code extension was compromised and installed backdoors on developer machines, with some observers asking whether the episode is connected to the recent Binance hack.

Pika H3 Flex generates faster than realtime.

Pika H3 Flex opens its weights, faster than realtime

Six denoising steps instead of fifty, priced from one cent per second.

Pika released H3 Flex as an open-weight model that generates faster than real time. The team post-trained it to run in six denoising steps rather than fifty, then refined the result for quality and creative control, with pricing starting at $0.01 per second.

HARDWARE · SPACE

A TPU in orbit answers its first prompt

Project Suncatcher hit a "hello world" milestone: a command was sent to orbit, the prototype satellite's TPU booted, ran a prompt, and returned a complete response — a haiku — back to Earth. Sundar Pichai framed it as an early step toward computing at the edge of the atmosphere.

MODELS · RESEARCH

Elsewhere on the frontier, Perplexity and Google DeepMind both claimed leadership positions for Gemini 4 Carbon in long-form research and on the Arena text leaderboard.

Day-0 RL for Mistral Large 4

On eight NVIDIA GB200 nodes, LoRA reinforcement learning on the NVFP4 checkpoint lifted AIME 2024 avg@8 from about 64% to about 75% in 75 training steps. vLLM generates the rollouts while FSDP2 or Megatron handles training.

Gemini 4 Carbon leads research

Perplexity's early evaluation calls Gemini 4 Carbon the strongest model seen so far for long-form research and answer quality, setting a new WANDR benchmark. It is coming to Pro and Max users inside Computer.

OpenAI launches Private Intelligence

OpenAI argues privacy should advance as fast as intelligence. Private Intelligence gives businesses stronger data controls while using the most capable models, built alongside some of OpenAI's largest customers.

THEORY · COMPLEXITY

Academics open a forum on OpenAI's ETH rebuttal

Impagliazzo, Paturi, and Williams set up a discussion page inviting peers to review OpenAI's claimed refutation of the Exponential Time Hypothesis. The authors say OpenAI notified them of the result and plans to release supporting material, and they are asking the community to weigh in.

"You're a hand-weaver in a world of power looms."

— ETHAN MOLLICK, ON CRAFT VERSUS INDUSTRIALIZATION

PRODUCTS & MODELS OCT · 31

PRODUCT

Runway Agent now runs files and code

It opens files, executes code, and builds launch kits, product shots, ad variations, and motion graphics from a reference.

PRODUCT

Kimi launches 3D Make

Kimi connects to Blender and CAD tools to create models, assemble scenes, and develop games.

INDUSTRY

Cohere doubles down on Canada

Plans to invest more than C$4 billion in R&D and AI infrastructure and create 1,500 jobs by 2031.

MODELS

Stable Audio 3.0 is fully licensed

Trained on licensed data, available through the API and as open weights on Hugging Face.

MODELS

Prism claims humanlike speech-to-speech

The first speech-to-speech model, it says, that is hard to tell from a human in conversation.

MODELS

Phonic-1 targets agentic speech

An advanced speech-to-speech model built for agentic tasks.

MODELS

Carbon-X 14B rivals GPT-6.1

Hugging Face researchers say the open 14B model competes with GPT-6.1 Sol on genomics benchmarks.

BENCHMARK

Gemini 4 Carbon tops Arena text

It lands first in Text Arena and high on several other leaderboards.

SECURITY

Accenture agent ran unauthorized

Reports allege a Claude AI Agent ran for two months inside a Pentagon security system.

SECURITY

Claude Code extension "compromised"

Amjad says an official extension was compromised and installed backdoors on developer machines.

MODELS

Microsoft ships a 4B computer-use agent

A hybrid policy on Hugging Face that decides when to click and act.

PAPER

Platonic representation, now a theorem

JEPA models trained separately on different modalities provably learn Platonic representations.

PAPER

A-RSI: zero-data multimodal self-improvement

It refreshes task pools when rewards plateau and self-generates questions from a program library; 800 new tasks lifted average performance from 57.77 to 61.05 across eleven benchmarks.

PAPER

LOWN routes vision only when needed

A training-free diffusion VLM that processes images only when text and image-conditioned predictions disagree, retaining 99.6% of LLaDA-V across ten benchmarks.

PAPER