Skip to content
Neural Daily

Articles

Every article we have published, newest first. Search titles and leads, or change the order.

Cursor logo in white on a black background

ToolsNews

Backdoored Claude Code extension hit Cursor's marketplace

For about 11 hours on October 29, a modified Anthropic.claude-code extension (2.1.312 and 2.1.313) that installs an SSH backdoor was available in Cursor's VS Code extension marketplace. If it updated in that window, check your machine.

· 3 min read

GitHub repository card for microsoft/TokenEconomizer with the Microsoft logo

ToolsNews

TokenEconomizer cuts coding agent costs by 18–31%

An open-source plugin from Microsoft Research and UC Berkeley learns from past sessions in a repository and wastes fewer tokens in the ones that follow. On SWE-Bench tasks it cut mean cost by 18.4% to 30.9%.

· 2 min read

OpenAI title card reading Community Safety Report and New Protections for Teens on a yellow-green gradient

PolicyNews

ChatGPT tightens violence safeguards and teen protections

OpenAI is tightening how ChatGPT handles violence-related requests: the model can now end a conversation on its own in some cases, and minors get less weapons-related information. Starting in early 2027, the company will publish a public community safety report every six months.

· 2 min read

Abstract illustration of staggered teal bars on a dark green background

ResearchNews

GLM 5.3 built Firefox exploits from the patch gap

In an Epoch AI experiment, the open-weight GLM 5.3 autonomously built working Firefox exploit chains from security fixes that were public but not yet shipped. Each successful attack cost about $1,000 in API fees.

· 2 min read

Line chart showing Samaya Sterling rising from 42.5% to 58.9% on FrontierFinance across training stages, above Claude Fable 5.1

ModelsNews

Samaya Sterling: 3B active params, beats Claude Fable 5.1

Samaya post-trained a 30B Nemotron model with 3B active parameters for financial research. It scores 58.9% on Samaya's FrontierFinance benchmark, above Claude Fable 5.1's 55.9%, at roughly an eleventh of the cost.

· 2 min read

Close-up of a bronze Lady Justice statue holding scales

PolicyNews

Wikimedia in NYT v OpenAI: training yes, RAG needs a license

The Wikimedia Foundation, which runs Wikipedia, filed an amicus brief in the New York Times copyright case against OpenAI. It argues LLM training can be fair use, but RAG that copies sources back into answers should need a license.

· 2 min read

Sebastian Thrun speaking on stage in a blue shirt

BusinessNews

Sebastian Thrun's Dulo raises $130M for hardware AI

Dulo, co-founded by Google X co-founder Sebastian Thrun and Adrian Aoun, raised $130 million to build foundation models for hardware design. Eric Schmidt, Marc Benioff and John Doerr joined the round.

· 2 min read

Three iPhones showing Apple Intelligence features with a Chinese-language interface

BusinessNews

Apple Intelligence is coming to China, with Qwen support

Apple will ship its first Apple Intelligence features in China in a developer beta from November 4 and in iOS 27.3 in December. Alibaba's Qwen will be available in some experiences, but the new AI-powered Siri is not included.

· 2 min read

Reddit Snoo logo on an orange background

BusinessNews

Reddit moves behind logins as AI search cuts traffic

Reddit is shifting to a primarily logged-in experience. AI-powered search is sharply cutting traffic from Google, and the company also wants to protect its data from AI scraping so it can license it to model developers instead.

· 3 min read

Aerial view of the Pentagon building in Arlington, Virginia

PolicyNews

WIRED: Contractor gave AI agent access to Pentagon system

According to WIRED's sources, an Accenture contractor connected a secure Pentagon system to their personal Claude Code account, and the agent ran on the network for two months. It appears to have sent login credentials and a network map to a high-risk third-party site. Accenture says the allegations are without merit.

· 3 min read

IBM Granite Switch illustration: a large modular cube with a green top and one block pulled out, beside a separate blue and a purple cube

ModelsNews

Granite Switch 4.1.1: update models like software

IBM released Granite Switch 4.1.1, a patch release of its modular model. The base model is untouched; only the decision adapter and the speech recognizer were updated, so integrations built on 4.1 keep working unchanged.

· 2 min read

Seal of the United States Department of Justice on a navy background

PolicyNews

DOJ is asking questions about Nvidia's Hugging Face deal

According to Reuters' sources, the US Justice Department has asked how Nvidia's $12.93 billion acquisition of Hugging Face would affect competition in AI. The questions are part of a years-long investigation into Nvidia's business conduct.

· 2 min read

Project ProtoStar announcement graphic from Hugging Face

ResearchNews

Project ProtoStar makes public science AI-ready

Hugging Face, US universities and nonprofit partners launched Project ProtoStar to release publicly funded datasets, models and code in AI-ready, reusable formats. The first open challenge starts November 2.

· 2 min read

How we detect frontier model misalignment cover graphic from OpenAI

ResearchNews

OpenAI details how it monitors models for misalignment

OpenAI described how it watches its agents' reasoning, actions and environment. After the Hugging Face incident, monitoring now covers training, evaluation, internal use and, in modified form, external deployment, with auto-pause and webhook alerts for Astra-class API use.

· 3 min read

OpenAI cover graphic with the title An updated Model Spec for more capable and beneficial AI on an orange and blue gradient

ResearchNews

OpenAI rebuilds its Model Spec around five core values

OpenAI updated the Model Spec, its public description of how its models should behave. The new version organizes guidance around five core values, adds sections on context, oversight and agency, and retires most of the detailed style guidance.

· 3 min read

OpenAI Developers card titled API credits for ChatGPT subscribers on a pink and orange gradient

BusinessNews

OpenAI doubles Pro Codex limits, adds monthly API credits

On October 29 OpenAI doubled the Work and Codex allowance on its Pro 100 and Pro 500 plans and canceled a planned cut on Pro 200. A day earlier, another perk went live: Pro and Business subscribers can claim up to $500 a month in API credits.

· 3 min read

Gemini file editing demo with a chart and a prompt bar reading show this as a graph instead

ToolsNews

Gemini now lets you edit generated files right in chat

The Gemini app can now preview and edit the documents, spreadsheets and slide decks it generates without leaving the chat. PDFs work on web and mobile, while other formats are web-only for now.

· 2 min read

Graphic announcing that Daily and Pipecat are joining Deepgram on a purple background

BusinessNews

Deepgram buys Daily, the company behind Pipecat

Voice AI infrastructure company Deepgram is acquiring Daily, maker of the popular open-source Pipecat voice agent framework. Pipecat stays open source under an independent foundation, and Deepgram also introduced a new open protocol, the Real-Time Context Standard.

· 3 min read