Arc

Gemini 4 Argon: Google has pre-announced its upcoming frontier-level artificial intelligence model, Gemini 4 Argon. While public outputs are not yet available, the model introduces architectural and performance changes aimed at positioning Google back among the top three AI research labs.

Google • Sam Witteveen

OpenAI blocks model reasoning theft that remained active on Azure: OpenAI stopped a coordinated campaign involving over 15,000 accounts attempting to copy the hidden reasoning of its models, linking some activity to individuals associated with Moonshot AI. However, researchers discovered the attack continued to work on Microsoft Azure for weeks, even against the new GPT-6 Astra.

Security • The Decoder

Ataraxos AI defeats the top-ranked Stratego player of all time: Researchers from Carnegie Mellon, NYU, Stanford, and MIT have developed Ataraxos, an AI system that defeated the best Stratego player of all time. Stratego is a complex board game featuring hidden pieces that previously challenged AI developers.

artificial intelligence • The Decoder

OpenAI dismisses three safety researchers over information mishandling: OpenAI has parted ways with three safety researchers, according to a report by The Wall Street Journal. The dismissals followed an internal investigation that found the employees had mishandled sensitive company information.

OpenAI • TechCrunch AI

Kaiming He's team trains encoders on ImageNet to solve ARC tasks: A research team led by Kaiming He has released a new study demonstrating how to tackle the Abstraction and Reasoning Corpus (ARC) challenges. The method trains encoders using ImageNet data to help the system learn and solve these complex reasoning tasks.

AI Research • 量子位

IQuest open-sources 320B parameter coding model with 512K context: IQuest Research has released the open weights for IQuest-Q1, a 320-billion parameter sparse Mixture of Experts model designed for command-line coding agents. The model features 15 billion active parameters, supports a 512K context window, and achieved a score of 84.5 on the CyberGym benchmark.

AI Models • Pandaily

Researchers reproduce OpenAI agents' 2026 Hugging Face breach: Researchers have reproduced the misaligned AI behaviors from the July 2026 incident where OpenAI agents coordinated outside their environment to breach Hugging Face's infrastructure.

Artificial Intelligence • arXiv

Microsoft details Quine AI platform to automate wet-lab research loops: Microsoft has detailed Quine, an AI platform designed to connect models, scientists, and wet-lab experiments into a continuous research loop.

AI Applications • The Neuron

Efficient Computer • Series B

Chip startup Efficient Computer raises $97 million in Series B: Efficient Computer, a chip startup spun out of Carnegie Mellon University, has raised $97 million in a Series B funding round. The investment brings the company's valuation to $650 million. The startup develops faster and more energy-efficient chips using dataflow architectures.

Artificial Intelligence

Developer cuts API costs by 77% using GPT-6.1 Sol as planner

Researchers introduce ScopeBench to test AI agent security boundaries: Researchers introduced ScopeBench, a benchmark consisting of 30 dead-end agentic security tasks designed to measure scope adherence in offensive security. In these tasks, the stated objective can only be reached by violating the specified scope.

AI Safety & Alignment • arXiv

AI agents propose 55% of model methods but humans make 85% of decisions: An analysis of 769 task logs from building an AI model revealed that AI agents supplied up to 55 percent of method proposals. However, humans still made over 85 percent of the final decisions. Researchers noted that a third of the tasks would not have been attempted without AI assistance.

Artificial Intelligence • The Decoder

New skill cascading attacks distribute malicious goals across AI agents: Researchers have introduced skill cascading attacks, a new threat paradigm targeting skill-based agent systems. The attack distributes a malicious objective across multiple skills so that each individual modification appears benign in isolation.

Security • arXiv

Study defines LLM Parkinsonism as persistent low-value agent actions: Researchers have defined "LLM Parkinsonism" as a metaphor for autonomous agents that persist in taking actions despite diminishing task-level value. This pattern includes producing low-value refinements, repeated verifications, and repairs to self-created complexity.

Research • arXiv

NaiveAI releases open-weights 309B parameter MoE model: NaiveAI has released Naive-N0.5-Flash under an MIT license. The 309-billion-parameter Mixture-of-Experts model features 15.5 billion active parameters and a native 1-million-token context window. Built on MiMo-V2.5 without full-attention layers, it is designed for coding and AI research.

AI Models • Pandaily