Redwood Research presents LinuxArena

- 20 live production environments for AI agents

- Frontier models achieve ~23% undetected sabotage vs. trusted monitors

- Useful work ≈ attack surface → sandboxing fails, monitoring is essential

Sakana AI discovered a prompting technique named String Seed of Thought (SSoT).

This work will be presented at ICLR2026

The method is simple: instruct the LLM to generate a random string in its own output, then manipulate that string to derive its answer.

It requires only a small addition to the prompt and no external random number generator a prompting technique named String Seed of Thought (SSoT).

SSoT significantly reduces output bias across a wide range of LLMs, both open and closed.

With reasoning models (such as DeepSeek-R1), it reaches accuracy close to that of actual random sampling.

The method generalizes from binary choices to n-way selections and arbitrary probability distributions.

On the NoveltyBench diversity benchmark, SSoT outperformed other approaches across all six categories while maintaining output quality.

AI4Science Catalyst backed by researchers from Stanford and Princeton unveiled LabWorld Factory, a "world engine" for biology.

The platform allows developers to generate fully scalable, simulated 3D biology labs entirely from natural language prompts.

Starting with a base of over 100 lab assets, the engine procedurally generates diverse layouts, physical tools, reagents, and camera views.

Cloudflare open-sourced an email client where an AI agent reads your inbox, drafts your replies, and never sends anything without your permission.

It's called Agentic Inbox. It runs entirely on Cloudflare Workers. Zero third-party servers touching your emails.

HuggingFace introduced ml-intern, the agent that just automated the post-training team

It's an open-source implementation of the real research loop that ML researchers do every day.

You give it a prompt, it researches papers, goes through citations, implements ideas in GPU sandboxes, iterates and builds deeply research-backed models for any use case. All built on the Hugging Face ecosystem.

How it works?

ml-intern makes full use of the HF ecosystem:
- finds papers on arxiv and hf.co/papers, reads them fully, walks citation graphs, pulls datasets referenced in methodology sections and on hf.co/datasets
- browses the Hub, reads recent docs, inspects datasets and reformats them before training so it doesn't waste GPU hours on bad data
- launches training jobs on HF Jobs if no local GPUs are available, monitors runs, reads its own eval outputs, diagnoses failures, retrains

ml-intern deeply embodies how researchers work and think. It knows how data should look like and what good models feel like.

CLI
Web + mobile

Also provisioned 1k$ GPU resources and Anthropic credits for you to use.

This is cool. Meet Simula from Google

A Google framework produces better synthetic datasets.

Instead of just randomly generating data or copying existing examples, Simula uses AI to strategically architect an entire dataset from scratch.

Anthropic's Mythos AI model is being accessed by unauthorized users 🤣

A handful of users in a private online forum gained access to Mythos on the same day that Anthropic first announced a plan to release the model to a limited number of companies for testing purposes, said the person, who asked not to be named for fear of reprisal. The group has been using Mythos regularly since then, though not for cybersecurity purposes, said the person, who corroborated the account with screenshots and a live demonstration of the model.

The users relied on a mix of tactics to get into Mythos. These included using access the person had as a worker at a third-party contractor for Anthropic and trying commonly used internet sleuthing tools often employed by cybersecurity researchers, the person said. The users are part of a private Discord channel that focuses on hunting for information about unreleased models, including by using bots to scour for details that Anthropic and others have posted on unsecured websites such as GitHub.

“We’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments,” a spokesperson for Anthropic said in a statement.

MIT & the IMO released MathNet, the world’s largest dataset of International Math Olympiad problems & solutions.

MathNet is 5x larger than previous datasets & is sourced from over 40 countries across 4 decades.

Lobster Capital became the first agent-ready VC.

Team published an llms.txt, a structured file that AI agents (ChatGPT, Claude, Perplexity) read to understand who they are, what they invest in, and how to reach their.

Why it matters: founders and LPs increasingly research funds through AI. The VCs who don't show up in those answers won't get the meeting.

Google launched Gemini Enterprise Agent Platform a platform for businesses to develop, scale, govern and optimize agents.

It’s the evolution of Vertex AI, bringing together model selection and agent building with new features for integration, security and more.

It gives access to 200+ of the world’s leading models through the Model Garden.

This includes Gemini 3.1 Pro, Gemini 3.1 Flash Image, and Lyria 3, alongside open models like Gemma 4.

Incredible work by Sony.They’ve built “Ace”, an autonomous ping-pong robot that uses RL and Sony’s vision sensors to achieve expert-level play in ping pong. A huge leap forward for adaptive robotics.

Ace robot beats 3 of 5 elite table tennis players. Loses to professionals.

Human players win points with faster-than-average shots (p<0.001 between won vs returned). Ace wins with ordinary shots. Same speed and spin profile whether it wins or loses the rally (p=0.88).

It's playing a completely different sport than the humans are.

Trained entirely in simulation. Zero sim-to-real tricks beyond good physics modeling and asymmetric actor-critic (critic sees ground truth, actor sees noisy sensors).

GitHub.

Autonomous AI agents just found 10 zero-days in Chrome and the model almost didn’t matter

AgentFlow, a multi-agent harness developed by researchers at UCSB and Fuzzland, autonomously discovered 10 previously unknown vulnerabilities in Google Chrome over 7 days — including 2 Critical sandbox-escape CVEs confirmed by Google’s Vulnerability Reward Program.

The model powering the campaign was Kimi K2.5 managed to find and exploit 10 vulnerabilities in browsers.

AgentFlow used Kimi K2.5 not because it’s the best model, but because running 192 parallel agents for 7 days at Claude Opus prices was impractical and found 10 zero-days in Chrome, including 2 Critical sandbox-escape CVEs confirmed by Google VRP.

Sandbox escape is a serious vulnerability class that breaks Chrome’s process isolation, but turning it into a full system compromise requires a multi-step exploit chain.

The paper withholds the PoCs entirely.

The real story here: the harness architecture did most of the work.

The same framework, with Claude Opus 4.6, scored #1 on TerminalBench-2. The model matters less than how you orchestrate it.

Google just proved Image generators are generalist vision learners

They introduced Vision Banana, a model built by instruction-tuning a base image generator (Nano Banana Pro).

Instead of using special systems for different tasks, they reframe every vision problem like segmentation or depth estimation as simply generating a new image. Think of it as drawing the answer instead of calculating it.

Vision Banana beats domain-specific experts, including the Segment Anything Model 3 (SAM 3) on segmentation and the Depth Anything series on metric depth estimation, all without sacrificing its original ability to create images.

This suggests generative pretraining is the new foundation for all of computer vision.

Anthropic just introduced forked subagents in their latest update

Unlike regular subagents, forked subagents can inherit the same context as the main agent. This looks convenient for cases where richer context matters more.

DeepSeek-V4 Preview is officially live & open-sourced

DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.

DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.

Open weights.

OMG 😯AI companies Cohere of Canada and Aleph Alpha of Germany have agreed to merge

Wow! Google presents a new Transformer alternative at #ICLR2026!

Meet MesaNet, proposing a new linear sequence layer that optimally learns in-context given a fixed memory budget.

Meet kUPS a molecular simulation engine built for the AI era, optimized for GPU in collaboration with NVIDIA.

kUPS is a plug-and-play, Python-native toolkit designed to integrate seamlessly with modern ML workflows.

GitHub.

Sakana AI shared a new work“TRINITY: An Evolved LLM Coordinator”

In nature, complex problems are rarely solved by a single monolithic entity, but rather by the coordinated efforts of specialized individuals working together. Yet, modern AI development is heavily focused on endlessly scaling up single, massive monolithic models, yielding diminishing returns. While model merging offers a way to combine different skills, it is often impractical due to mismatched neural architectures and the closed-source nature of top-performing models.

TRINITY is a system that fuses the complementary strengths of diverse, SOTA models without needing to modify their underlying weights.

TRINITY processes queries over multiple turns. At each step, a lightweight coordinator assigns one of three distinct roles to an LLM from its available pool:

1. Thinker: Devises high-level strategies and analyzes the current state.

2. Worker: Executes concrete problem-solving steps.

3. Verifier: Evaluates if the current solution is complete and correct.

By dynamically assigning these roles, the coordinator effectively offloads complex reasoning and skill execution onto the external models.

What makes TRINITY unique is its extreme efficiency. The coordinator relies on the hidden states of a compact language model and a small routing head. In total, it has fewer than 20K learnable parameters.

Training this system presented a massive challenge. Traditional Reinforcement Learning (REINFORCE) failed because the gradients had a low signal-to-noise ratio due to binary rewards and weak parameter coupling. Imitation learning (Supervised Fine-Tuning) was ruled out because generating multi-turn labels is prohibitively expensive.

Sakana turned to nature-inspired algorithms. Team optimized the coordinator using a derivative-free evolutionary algorithm. Sakana found that evolution is uniquely suited to optimize this tight, high-dimensional coordination problem where traditional gradient-based methods fail.

The results are very promising. In experiments, TRINITY consistently outperforms existing multi-agent methods and individual models across various benchmarks.

More importantly, it demonstrated incredible generalization. Without any retraining, TRINITY transferred zero-shot to four unseen tasks (AIME, BigCodeBench, MT-Bench, and GPQA). On average, the evolved coordinator surpassed every individual constituent model in its pool, including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.

This work is central to Sakana AI's vision.

This foundational research is part of the core engine powering Sakana’s multi-agent product: Sakana Fugu.

This work instead describes the future of critical open world evaluations.

Most agentic benchmarks center around tasks that are automatically verifiable.

Benchmarks are saturated more quickly than ever. How should frontier AI evaluations evolve?

Research team argue that the AI community is already converging on an answer: Open-world evaluations. They are long, messy, real-world tasks that would be impractical for benchmarks.

A new project, CRUX, will run open-world evaluations regularly.

In a first experiment, team tasked an agent with publishing an iOS app to the App Store. The agent made two errors, but eventually succeeded after just one manual intervention.

Team released over 1GB logs from this experiment. Also pre-processed them for easy access in Docent.