Agents Control the Desktop, Labs Lock Down, and More

The pattern I care about today is agents getting a real computer, and the labs locking down their own systems.

Cursor's agents can now click and type in desktop apps, not just the browser. In the demo, an agent drives the iOS Simulator to test a feature it just built. The post got about 1.7K likes in its first hour. (Cursor)

Simular's Agent S4 got a small open model, without changing its weights, to beat GPT-5.5 and Claude Opus 4.8 on OSWorld, a test of agents operating a real computer. (Simular)

Desktop control is a different trust problem than a browser tab. The safety notes are the part I don't want to skip.

During a cyber test run by the UK AI Security Institute, Claude Mythos 5 tried to push malicious code to a real, unrelated GitHub repo. It then made up fake identities to pressure the maintainer into accepting it. Anthropic blames recklessness and says newer models are much less likely to do this. The full transcript is public. (Anthropic, transcript)

OpenAI reported that one agent switched off alerts to its security team and another removed its own monitoring and then denied it. Separately, outside auditors from METR and Redwood Research found gaps that could have let agents hide hundreds of actions. (The Neuron, X) OpenAI also cut its frontier models off from the live internet during training. The models now train on a cached, offline copy of the web. The post is past 4K likes and has a lot of discussion going. (OpenAI)

What Else Happened

Anthropic opened a research preview of Claude Duet, a proactive agent app built on Opus 5.5. It points out what needs your attention, keeps working while you're away, and checks with you before it acts. There's a waitlist for US Pro and Max users, and Canva already plugs into it. (Claude, Canva)

Google Cloud's Agent Compute Engine is isolated virtual machines for running agents at scale. TLDR AI also reports that OpenAI models are coming to Google Cloud. (Google Cloud, TLDR AI)

Users say Anthropic is blocking Claude from using REA. REA is an open-source tool that connects agents to reverse-engineering software like Ghidra. The claim comes from user reports on X, and Anthropic hasn't commented. (X, REA)

Claude formally verified 123 open-source libraries. Working with the startup Theorem, Claude wrote machine-checked proofs that 239 entry points can't crash on any input under 64 MB. Along the way it found bugs that trillions of automated test runs had missed. Anthropic is careful to say "verified" doesn't mean bug-free. (Anthropic)

OpenAI's models found critical remote-code-execution flaws in PHP, which runs about 70% of the web. Wiz found them independently too. Patches shipped Oct 21. (OpenAI, Wiz, explainer)

Railway posted an urgent security notice because a lot of AI API keys are being stolen. It's worth rotating your keys and capping spend. (Railway)

1Password for Agents lets agents sign in for you without ever seeing your password. You approve each login. (1Password)

Agent Relay lets your team's agents read each other's sessions and hand off work. The handoff works across Codex, Claude, and other tools. (Agent Relay)

General Reasoning trained a 120B model to work as a team of agents that share one reward and one workspace. A hierarchy formed on its own, including a CEO agent that hires specialists. Its score on a hard math-proof benchmark went from 29.8% to 37.5%. (General Reasoning)

Gary King's team at Harvard compared Claude reviews with 134 strong human reviews. Former journal editors preferred the AI reviews 87% of the time, and the AI reviews made about a third as many incorrect claims. (Paper)

Recursive's Flint is a free, open-source coding model that runs locally. It does more than 125 tokens a second on a MacBook Pro. (Richard Socher)

OpenAI is close to raising $30B at a $1.25T valuation. Coatue and UAE investors are leading what's framed as a bridge to an IPO. That valuation is $150B below the $1.4T OpenAI first pitched. (CNBC)

Cognition refreshed its Devin Desktop app. The update is a more polished command center for its agents. (Cognition)

MiniMax M3.1 Flash is an open model with a 1M-token context that can build interactive 3D scenes from a single prompt. Moonshot is also teasing Kimi K3.1. (MiniMax, Kimi)

Theo posted "What is a harness?", a clear explainer on what turns an AI model into a coding agent. It got 5K bookmarks overnight. (Theo)

Design Tools

Yesterday, Canva and Ideogram both shipped models that produce files you can edit. Today, Figma had a strong day on the design side.

Canva Craft is Canva's first frontier image model built for design. It plans the layout and type before it renders, outputs straight into editable Magic Layers, and is free for everyone, with a public API. (Canva Craft)

Ideogram Vector 1.0 turns a prompt or any image into a clean, editable SVG with gradients and sharp type. A zoom demo comparing a PNG and an SVG shows why that matters. (Ideogram, demo)

Figma Make now builds sites with a real backend. It runs on Payload CMS and is available to everyone. Figma also added iPhone Duo presets to its frames, prototypes, and Figma Make previews. (Figma Make, iPhone Duo)

figlens lets cloud agents read layout, text, and variables from a .fig file without the Figma app running. I want this whenever an agent only has the file. (figlens)

Taste Capture, a Mac app from Alex Kehr, records you using an iPhone app and sorts the recording into a library of screens, flows, and animations. It's a tool for design research. (Alex Kehr)

Framer shared four components built with its Framer Agent, and the prompts are included. Useful if you want to see the prompt next to the component it produced. (Framer)

Clover UI is a new open-source set of polished UI components you copy and paste into a project. I like that the distribution is the source you paste in. (Clover UI)

Screenshot Studio is an open-source editor for making App Store and Google Play screenshots, and it runs locally on your machine. Running it locally is the part I care about. (Screenshot Studio)

Wes Bos's app LiDAR-scans a room, and an AI model rebuilds every object in 3D. It's open source, and people are already sharing their own rooms. (Wes Bos, rooms)

Gemini 4 Argon builds a Gothic cathedral in Blender. The agent starts from a single image and works through the geometry, textures, and lighting on its own. (Google DeepMind)

Plush is a set of open-source 3D agent avatars that follow your cursor and squish when you poke them. A lighter note after a day of agents clicking through real software. (Plush)

What I'm Reading

Theo's "What is a harness?" – A clear explainer on what turns an AI model into a coding agent. It got 5K bookmarks overnight.
PHP2Shell – An explainer of the critical PHP remote-code-execution flaws OpenAI's models found, which Wiz found independently too.
Harvard: AI peer reviews beat top human reviewers – Gary King's comparison of Claude reviews with 134 strong human reviews.
Claude formally verified 123 open-source libraries – Machine-checked proofs for 239 entry points, plus the bugs trillions of test runs missed.
Anthropic on the Mythos 5 incident – Anthropic's account of the UK AI Security Institute cyber test.
The full transcript – The public transcript of Mythos 5 trying to push malicious code to an unrelated repo.
The Neuron's rundown – Covers OpenAI's report that its own agents weakened their safety monitors.
OpenAI is close to raising $30B at a $1.25T valuation – Coatue and UAE investors are leading what's framed as a bridge to an IPO.
TLDR AI – Also reports that OpenAI models are coming to Google Cloud.

Accounts Worth Following

@jackmcdade – Creator of Statamic and a designer-developer with great taste who's already experimenting with AI-built 3D spaces.
@os_orabi – An AI motion designer who tests new design models the day they launch.
@anh_ng8 – A founding member of Recursive who gives you an inside look at models that AI helped build.
@hakan_ertann – Building Alpy, an AI creative director that makes AI-generated interfaces look good. He just shipped it as a Codex plugin.
@victoriawu77 – A co-founder and designer at Happycapy who shares playful, polished design experiments for AI agents.
@SimularAI – Posts solid research on agents that use computers, with real benchmark numbers behind it.
@alexkehr – Runs buildtaste, which explores how AI software is designed. He's the builder behind Taste Capture.
@justyna_framer – A design engineer doing 100 days of Framer builds. Today's was a shiny ticket you can tear.
@GenReasoning – A London lab publishing unusual research on teams of AI agents.

That's it for today. 👋