Anthropic opened a research preview of Claude Duet, a new app where Claude works as a proactive agent

It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

Anthropic designed Claude Duet around Claude Opus 5.5. It carries over the memory, skills, and connected apps from your Claude account, so there’s little to set up.

Waitlist.

The China state-aligned threat actor Storm-0940 used the agentic AI penetration testing framework ARTEX to automate attacks targeting hundreds of organizations globally across a broad range of industries.

Microsoft Threat Intelligence assesses that Storm-0940 was responsible for the recently disclosed ConnectWise supply chain compromise of a multitenant application.

AI-accelerated follow-on activity included reconnaissance, application credential changes, and attempts to modify privileges that could enable expanded or persistent access.

Dream Labs released Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.

Try it here.

Anthropic presented Agent4Sci, which gives frontier coding agents 20 scientific ML benchmarks across 11 domains and asks them to beat the published best model.

They beat it on all 20. Claude Opus improves on the published best by 10.4% on average, and no run was flagged for misconduct.

The gains come almost entirely from standard ML-engineering tricks. The trajectories show agents misjudging what is novel, leaving compute unused, and settling into small local sweeps.

Adding three lightweight skills for research ideation, experiment management, and scheduled checklists raises Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, with more varied and more domain-relevant changes.

Moonshot Ai introduced K3.1 Pro

The first model in the K3.1 series is a 3 trillion parameter class model with a 1 million token context window.

Kimi is also introduced K3.1 Pro Highspeed, which it describes as offering the same intelligence at roughly 2–3× the output speed.

Open weights by Nov 4, 2026.

Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.

This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.

The emerging rules of thumb:

1. Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.

2. Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.

3. Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.

The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.

Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.

As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.

Anthropic has filed for an IPO with the SEC

Its first highlighted potential use of AI: "profoundly accelerate the rate of progress in biology and medicine, potentially curing diseases that have plagued humanity for millennia”.

And here’s what the filing says.

Anthropic says it has grown from zero to tens of billions of dollars in 3 years:
2023: $41M. 2025: $4.6B.

In the first half of 2026 alone, revenue reached $16.3B, roughly 14x more than a year earlier.

Anthropic has more than 1,900 customers paying over $1M a year, and around 250 paying over $10M.

Existing customers increase their spending roughly 9x over the course of a year.

According to the prospectus, the business is close to operating breakeven.

Adjusted operating margin for the half-year came in at +10%.
At the same time, net loss stands at $80.7B.

They explain that the more Anthropic is worth, the bigger its reported loss.

Anthropic’s main bet is cognitive work. The company values the work of the world’s knowledge workers at $40 trillion a year.

Their goal is to turn AI from an assistant into a Virtual Collaborator: an agent that independently carries out tasks inside an organization for days at a time.

The path there is laid out like this: first, developers move to AI through Claude Code, and then employees across every other profession hand their work over to AI through Cowork.

Their next bet is science and the physical world. At the top of that list are biology and drug discovery, followed by robotics, manufacturing, energy, and materials science.

The most unusual part of the prospectus is the governance structure.

Anthropic is incorporated as a Public Benefit Corporation. The majority of the board is elected by the Long-Term Benefit Trust, an independent trust with no financial stake in the company.

The 7 co-founders will receive a single share that carries 50.1% of the voting power and no economic rights.

OpenAI just shook one of the foundations of computer science.

In a new paper bluntly titled “The Exponential Time Hypothesis Is False,” OpenAI claims a deterministic, subexponential-time algorithm for exactly counting solutions to k-SAT which would refute ETH, one of the central assumptions underpinning modern fine-grained complexity theory.

ETH has long been used to argue that huge families of problems probably cannot be solved substantially faster. If this proof holds, a large body of conditional lower bounds built on ETH and SETH will need to be re-examined.

The paper also claims consequences for SAT, graph coloring, independent sets, perfect matchings, and more.

Google DeepMind researchers have developed AudioFM, a self-supervised framework for general-purpose audio representations

How it works: a student sees time- or frequency-masked spectrograms, predicts the masked patch representations, and has to pick the matching full-view EMA-teacher target from candidates in other recordings. Plus weaker supervision from masked temporal neighbors.

Results: on linear probing it achieves the best or joint-best performance on 10 of 12 tasks.

Something particularly interesting: beats prior audio SSL on heart murmur (+4.7 F1) & sleep apnea subtype (+3.2 F1) classification.

The largest math proof in human history is now computer-verified.

Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry."

About 15,000 pages, built by over 100 mathematicians across generations.

AxiomProver's swarm built the proof in 50 million lines of Lean, over 24 days.

The repository is the largest Lean code base to date.

Code generation at this scale forced us to reimagine the underlying engineering infrastructure.

The largest sporadic group, the Monster, has ~8 × 10⁵³ elements.

Related to it is the "Monstrous Moonshine" conjecture, Richard Borcherds’ 1998 Fields Medal work.

Yet for long, there is cloud on the correctness of the proof, as no one person has read the whole - it’s too big.

CFSG runs through the Unbalanced Group Theorem (UGT).

Axiom’s AI found counterexamples to a claim in the published proof.

Fortunately, UGT holds.

AxiomProver proved a new result to replace the step, built on ideas from Ronald Solomon.

It also fixed hundreds of major errata, TBA.

New from Anthropic: How AI agents shape experts' judgment calls

When experts delegate work to an agent, do they still make the key decisions?

Researchers tested this in a randomized trial with 50 professional data analysts.

In experiment, analysts used Claude Code to advise a fictional farming co-op on whether to expand a program.

The analysis had four decisions, each with two defensible options. Team randomly assigned which ones the agent would lean toward.

AI is cracking many famous conjectures. But can it ask new questions worth answering?

With a team of mathematicians and physicists, researchers from Stanford built After42, a system that generates novel research problems that human experts find genuinely interesting.

The key idea: develop questions by trying to solve them.

Team also asked whether AI can judge which questions are interesting. Researchers find that model judges' top choices align with expert consensus but do not consistently capture individual experts' preferences.

after42.org

It’s a big! Formation bio announced the $600M Drug Development Accelerator Fund.

A new Fund brings together dedicated capital, drug development expertise, and AI-powered development capabilities to acquire promising drugs and advance them through the clinic.

Meet Lor.

Mirendil building Lor to be a frontier AI lab for every scientist.

Today, its agents write code, run experiments, and iterate on the results for days, on any machine, hundreds at a time.

What you can hand off to Lor:

1. Run these 12 ablations overnight & write a report.

2. Audit every file in this repo for security bugs & open a PR for each fix.

Run hundreds of agents at once.
In one run, 102 agents in Lor built environments from 100 PyTorch PRs, averaging 8 minutes each.

Lor comes with its own compute.

Agents request nodes, set them up, run the work, and release them when they’re done. No cloud account. No setup.

Start an experiment, a training run, or an eval sweep. The agent brings up the machines and hands them back when results are in.

In one overnight run, agents in Lor found security bugs in LiteLLM and fixed them by morning.

Proteins, materials, medical scans: more and more of science runs on AI. Building it takes a team.

It doesn’t have to. One Mirendil researcher just pretrained Mirendil-Base in 3.5 months

Lor is in private beta: free during the beta, on macOS, Linux, and web, running Claude, GPT-6, Gemini, open models, and own Fern.

The waitlist

Samaya introduced Sterling, a new model post-trained for challenging financial research that outperforms frontier models.

Researchers were able to post-train a much smaller model (Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters) to greater accuracy (58.9% vs 55.9% on FrontierFinance) and at 11x lower cost ($0.61 vs $6.89) than Claude Fable 5.1, the best-performing frontier model.

The team did this via a four-stage training process: two stages of SFT, where the model first learns to research with its own financial tools and how to construct a good agentic research strategy.

The remaining two stages are RL, with an offline GRPO stage where the model further hones its research strategy from eight teachers that provide diverse approaches, and an online GRPO stage where the model can make end-to-end attempts and learn from its mistakes.

Researchers see substantial gains from the co-adaptation of the model and the harness through post-training, both through the model learning to effectively use the high-quality financial tools and through better compaction approaches that let even a small model execute longer and deeper research.

The team is exploring multiple follow-up directions, from multi-agent research swarms using this small model, to training firm-specific models to encode unique judgment, to training a larger model, where early results show even greater gains from the training process.

Google announced something even bigger for Search Marketers

You can now build apps (using MCP) that can be used directly inside Google Search and Gemini.

Google announced the Connected Apps Platform.

When approved, users link your app once and can access their data and take actions in your service seamlessly from Google's products.

Cool paper from Amazon and colleagues on self-improving multi-agent systems.

Episodic memory for agents is so underrated.

They built EMMA a system that has three parts, all run by the same LLM.

A controller plans each task and runs it one step at a time, choosing which sub-agent and which skills to use at each step.

Sub-agents start as a generalist, and each specialist inherits the generalist's prompt plus its own role-specific method. Skills are short procedures, such as algorithms or pitfall checklists.

EMMA treats the controller's routing memory, the agent prompts, and the skills as episodic memory that improves from failures. After a failed task, a reflector decides which of the three owns the fix and proposes one edit there. Guards keep every accepted edit scoped, incremental, and bounded.

Cursor identified a security incident affecting a modified Claude Code VS Code extension accessed through Cursor IDE extension marketplace, impacting at most 5,000 users.

This malicious version of the extension was not developed by Anthropic. It does not impact other extensions downloaded from the cursor.com/marketplace, Grok Bot, Grok or our other SpaceXSI products.

In response, Cursor took down extension marketplace this afternoon, worked with relevant providers to neutralize the malware, and have reinstated the marketplace after confirming there is no further risk.

If affected, please review instructions in this blog post to understand how to find the malicious package and establish a path to mitigate any potential impact.

Google introduced a sleep framework in continual learning that enables LLMs to continuously integrate new facts and enhance reasoning while being robust to catastrophic forgetting.

Google proposed sleep cycle consists of two phases:

1. Memory consolidation: the model uses imitation learning and RL to perform upward distillation (called knowledge seeding) and transfer knowledge from fragile short-term parameters to persistent slow memories.

2. Dreaming: the model generates diverse synthetic data to improve its own performance.

Meet RT-J a System 1 model for Data Science: put raw data in context, predict in one forward pass.

No feature engineering, no model training, just accurate predictions.