OpenAI agent during evaluation, escaped sandboxing and hacked into HuggingFace.

Because OpenAI models don’t allow advanced cyber capabilities, HuggingFace used a Chinese open model to contain the rogue OpenAI agent.

Claude Code on desktop now works with the iOS simulator.

Build and run your iOS app, and the simulator opens in a panel right next to your conversation.

Claude can see the app as it runs, interact with it, and iterate until finished. Tap through the simulator yourself at any time.

Available in Claude Code on desktop (macOS only, requires Xcode).

Meet the world's fastest tokenizer implementation, Gigatoken

Gigatoken can tokenize the entire internet in <7 hours on one machine.

Specifically, this refers to tokenizing Common Crawl on a dual socket AMD Epyc server, reaching 24GB/s of throughput.

Gigatoken is ~500-1000x faster than HuggingFace, and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.

You can now ask Claude about the Anthropic Economic Index, our public dataset measuring how AI is used across the economy.

Ask which occupations use AI the most, or what kinds of tasks people are automating, and the answers draw directly from the Index data.

To turn it on, open the connectors menu and find the Anthropic Economic Index in the directory. It works in any conversation, and the full datasets remain free to download on our website.

YC dropped their Fall 2026 Requests for Startups and it’s one of the most ambitious lists yet.

YC wants founders rebuilding real world systems: education, defense, healthcare, infrastructure, work, and more.

Highlights from the new RFS:

1. The Primer an AI tutor that adapts like a private mentor, teaching young kids to read, write, and think at the level of Aristotle tutoring Alexander.

2. The Future of Defense. They’re looking for low-cost interceptors, next-gen sensors, drones, resilient logistics, and modular hardware that can survive extreme environments.

3. A Cloud for Small Software a new cloud built for single user or tiny team tools that agents can now generate easily. Think Google Docs-level sharing, but for custom software.

4. Multiplayer AI. Turning solo AI agents into team tools. Real time collaboration where multiple people + agents work together on long-running tasks (coding, deals, support, contracts, etc.).

5. Compute at Sea. Floating data centers on the ocean to solve power, land, and permitting problems.

6. AI-Powered Consumer Products for 1 Billion People. The next consumer giants built on AI. Intelligence is finally good enough and about to get cheap enough.

7. AI for the Aging Population. Tools for the massive demographic wave of seniors: conversational voice interfaces, monitoring, home robotics, and caregiver coordination.

8. New Operating Systems for the Physical World. Software for industries where 80% of workers aren’t at desks (construction, maintenance, fleet ops). Managing humans + robots + AI agents together.

9. Plus strong takes on Crypto (bear market = best time to build), Data for the Real World, and more.

These are not requirements. They’re signals of where they see massive opportunity. The best ideas often surprise them.

If you’re building in any of these spaces (or something even bolder), this is a great time to apply.

Andrew Ng announced OpenWorker an open-source agent that doesn't just chat with you, but delivers finished work like hand you a polished document, send a slack message, or update a calendar entry.

Ask it to prepare a customer brief, untangle your calendar, draft a report, or triage a Slack alert.

It works across your files and everyday tools, produces the deliverable, and checks in before doing anything consequential.

OpenWorker runs on your Mac, with Windows support coming soon.

It does not lock you into any one model. Bring your own API key and run it with GPT 5.6 Sol, Claude Fable, Gemini 3.6, an open weight model (like Kimi, GLM, DeepSeek, Inkling), or Ollama to keep your data local.

Your data does not leave your machine except through an LLM provider and integrations that you choose.

GitHub.

Tencent released a great paper on self-improving agent harnesses.

If you maintain a production agent harness, finding every file behind one behavior is often harder than writing the edit.

Harness Handbook builds a three-level map from runtime behaviors to source locations using static analysis and LLM-assisted structuring.

Its BGPD workflow guides coding agents from the system overview to relevant stages, functions, and files, then verifies every candidate against current source.

Anthropic Introduced Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

Opus 5 is stronger than Opus 4.8 on cybersecurity tasks. But it remains substantially behind Mythos 5 at developing exploits.

Its safeguards are designed to allow developers to identify and fix software vulnerabilities, while blocking high-risk uses.

Induction introduced imagination models: a new foundation model architecture that unlocks learning from internet-scale video.

The first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels.

Imagination models predict what happens next in a video in representation space.

Predicting future states teaches Photon-1 to use a computer, implicitly. After a small finetune to teach it to use the keyboard and mouse, Photon-1 can imagine the future and act to get there.

Despite only ever seeing computer video, Photon-1 learns general world concepts.

With a small amount of finetuning, it can predict billiard physics and checkers states better than baselines trained on the same data.

Photon-1 also picked up human habits from its pretraining video.

It learned to use AI tools the way people do. Sometimes it prompts ChatGPT, checks the output, and steers the LLM until the task is done.

Great paper from Google.
Why context beats scale for agents working against unfamiliar APIs.

GPU kernel optimization has KernelBench to hillclimb on. TPUs had nothing, and the Pallas DSL is documented thinly enough that models mostly guess.

New research from Google, Harvard, and UC Berkeley introduces JAXBench, 50 JAX workloads built from real MaxText architectures like Llama-3.1, DeepSeek-V3, Mixtral, Mamba-2, and AlphaFold2.

Eight operators ship with hand-tuned Pallas kernels from Tokamax, so agent output gets measured against expert work instead of a naive baseline.

With Gemini 3 Flash, conditioning on curated TPU documentation lifts per-sample correctness from 5.8% to 37.3%, solving 48 of 50 benchmarks at a 1.28x geomean speedup. Beam search then pushes it to 1.36x.

Correctness turned out to be a documentation problem and speed turned out to be a search problem.

Circle has acquired fundamental assets from the IBM blockchain patent portfolio, including 680+ patent families and nearly 1,000 issued patents worldwide.

The acquisition makes Circle the leading U.S. blockchain patent holder and strengthens the foundation behind USDC, CPN, Arc, and onchain financial infrastructure.

Moonshot released the model weights and technical report of Kimi K3.

Meet Waddle Labs: Claude Code for robots

Connect API to your robot and enter a prompt, then our agents write code to achieve the task in 20 minutes.

Waddle's agents enable new forms scaling.

Agent first wrote the “fold_grasp” skill while flipping a package; another agent later used it to fold a shirt. Over time, agents accumulate a library of skills, enabling future agents to compose them and achieve harder tasks.

Chinese AI companies may increasingly adopt a ‘paid weights’ commercial licensing strategy to monetise their open models, according to Goldman Sachs Hong Kong managing director Ronald Keung.

Note: Interview was done before yesterday's K3 weights drop with a modified license - so prediction proved quite prescient

The reasoning is that a shortage in compute forced Chinese companies to open-source their models, as this allowed them to build up their global brands while handing compute costs off to third-party inference providers.

Now that the companies are more established globally, they will increasingly want a bigger share of the revenue pie, Keung says.

New research from Meta and CMU. This one is on agentic context management for long horizon tasks.

Production agents accumulate context every turn. The usual fix compresses on a token threshold and throws the remainder away, so the trigger sometimes fires for reasons unrelated to what the agent is currently working on.

ACM hands the agent purpose-built context editing tools. It decides when to compress, offloads what it drops into an external memory store, and queries that store later when it needs the detail back. Compression shifts from short-term to long-term memory.

Code, data, and checkpoints are released.

Anthropic released MCP 2026-07-28, it's the largest update to the protocol since launch

MCP is now stateless, making it easier to deploy and scale remote servers.

Before, running a remote MCP server meant managing session state, which limited where you could run it.

Now that MCP is stateless, you can deploy on serverless and edge infrastructure, or scale horizontally behind any load balancer.

Extensions are also first-class a formal path to extend the protocol. Examples:

1. MCP Apps: server-rendered UIs in a sandboxed iframe

2. Tasks: long-running and async operations

3. Enterprise Managed Auth: control MCP server access centrally via your identity provider.

Also in this release: auth hardening and a formal deprecation policy.

OpenAI, Anthropic, Google DeepMind staff are circulating a letter asking the US government to support a mechanism that could help “deliberately pace” AI development if needed, bc of risks of the technology becoming out of control.

Why do AI-generated videos often look good but break physics?

Google researchers with University of Copenhagen and Oxford introduced VGGRPO a new method that trains a latent geometry model to directly read scene depth and motion from a video diffusion model’s internal representations.

Then it uses reinforcement learning to reward smooth camera paths and consistent 3D geometry.

Result: VGGRPO outperforms prior approaches in camera stability, geometry consistency, and overall quality while eliminating expensive VAE decoding.

Sakana AI presented Dreaming in Voxels: How AI is Generating Playable Minecraft Worlds

Generative AI has conquered images, video, text. But what about interactive 3D environments?

Sakana trained models on billions of cubes to generate fully playable, structured worlds.

Perplexity open-sourcing Numbat, an agent-detection and response layer that is designed to work across agent harnesses

Numbat works across agent harnesses on desktop, CLI, IDE, and gateway agents.

It supports live monitoring, local detection, pre-action blocking, and forensic reconstruction from existing session artifacts.

GitHub.