Malicious Claude Code Extension, Canva's Brush, and More
The story to act on today is a security one. Cursor found a modified, non-Anthropic version of the Claude Code VS Code extension in its extension marketplace. Cursor says it affected at most 5,000 users. It took the marketplace down, worked with providers to neutralize the malware, and then brought it back.
If you use Claude Code for VS Code, update it to 2.1.314. It's another reminder that editor extensions are now part of your attack surface, especially the ones that run agents with access to your machine. Cursor on X, security notice.
What Else Happened
Canva launched Brush, its own vision-language model for design. Brush reads a design's layout, layers, and style, then edits it in place, whether that's fixing one line, restyling a deck, or building from a blank page. Canva says edits are up to 69% faster than with its previous agent. It's Canva's second in-house model in about a week. (Canva on X)
Figma explained how it trained its own Figma-fluent editing model. The research post covers the first-party model behind the Figma agent's edits: fine-tuning on 250K+ synthetic edit examples in Figma's native format, now moving to reinforcement learning with rewards shaped by expert designers. The agent pairs this model for precise edits with third-party models for images, vectors, and new explorations. It's worth reading next to Canva's Brush, since both companies landed on the same idea: a model that speaks the tool's native format. (Figma's article on X)
OpenAI rewrote its Model Spec and detailed how it monitors frontier models. The updated Model Spec is built around five values (faithfulness to human intent, minimizing harm, honesty, human agency, and care) to guide judgment as models do more autonomous work. Separately, OpenAI said it now monitors all frontier RL runs, evals, and internal Codex use. In September, its monitors caught a model using a DNS misconfiguration to reach the live internet within 15 minutes. (Model Spec thread, updated Model Spec, monitoring post, detecting misalignment)
Anthropic trained Claude to flag misuse itself. With "self-assessment," Claude gives a quick safety read on the conversation right after it answers. Anthropic says this has caught misuse its other safeguards weren't built for, and that it gets more accurate as models improve. Anthropic also widened its AI for Science program to foundations and institutions, including the NSF and the Simons Foundation. (Anthropic on X, safeguards blog, AI for Science)
Thariq and DHH took opposite sides on agent workflows. Thariq, who recently left Claude Code, wrote that coding agents can create a kind of mania where you feel busy but nothing really ships, often as a form of avoidance. The post drew 12K+ likes. DHH pushed the other way, saying the real "AI psychosis" is insisting that humans review every line. Both threads are full of practitioner replies worth skimming. (Thariq's post, DHH's post)
Launches and Demos
- Factory Droid 1.0 – a new engine for Factory's coding agent. Programmatic tool calling is now the default, so the agent writes a program that can make many tool calls in one model request. It's on desktop, CLI, and SDK. (Matan Grinberg on Terminal-Bench)
- GitHub Copilot Fusion (GA) – announced at GitHub Universe. Pick "Copilot Fusion" and it routes and chains models per task, at Standard, High, and Max levels. Artificial Analysis puts Fusion Max within about 2 points of Claude Code on Opus 5.5 at about half the cost per task.
- Mistral "Le Chonk" (Mistral Large 4) – an open-weights release aimed at agentic coding and cybersecurity, ranked #1 among open models on Cognition's FrontierCode.
- herdr desktop app and herdr-core – an open-source (Apache 2.0) agent and terminal workspace where every part of the app is a file your agent can change. Terminals and agents keep running through app crashes and updates. (Atuin Threads mod)
- Antigravity plugins marketplace – search and install packaged customizations (tools, skills, workflows) in Google's agent IDE.
- Jules is winding down – Google is folding its async coding agent into Antigravity's background tasks.
- Cartesia Managed Agents – an end-to-end enterprise voice-agent platform, #1 on Artificial Analysis Speech Agent Arena and ServiceNow EVA-Bench.
- Claude mobile on iPad and iPhone Duo – what Claude makes now opens beside the chat instead of over it, and can go full screen.
- Framer MCP for migrations and on-the-go fixes – use any AI agent over MCP to rebuild pages, migrate CMS content, and ship site fixes without a laptop.
- DesignMD Remix – take any design system, change its colors, fonts, corners, spacing, and shadows with a live light and dark preview, then export your own DESIGN.md for agents.
Accounts Worth Following
- Karl (@_kejk) – Founding Design Engineer at DuckDuckGo, with sharp takes on what design engineering means now that "prompting Claude" is common. (example)
- JP (@dqnamo) – product engineer at Interface London who posts polished, craft-focused UI and interaction work.
- herdr (@herdrdev) – an open-source agent and terminal workspace that's shipping fast. Good for watching where agent dev environments are going.
What I'm Reading
- DeepMind's "Organizing Intelligence" – why AI's economic impact may depend more on how well agents coordinate in teams than on solo skill (Imas, Koh, Legg)
- Companies pausing aggressive AI rollouts – Meta, Microsoft, and Walmart pulling back after errors and cleanup costs
- Chinese labs reportedly distilling GPT-6 Astra – Kimi K3.1 Pro and DeepSeek V4.1 Flash were reportedly trained on Astra outputs after Anthropic cut off Claude distillation
That's it for today. 👋