The stories that matter
Top line
Story 1 of 3
A malicious Claude Code extension was served from Cursor's extension marketplace
A modified
Anthropic.claude-code2.1.313, not made by Anthropic, was available from about 4:27 AM to 3:35 PM CT on Thursday, Oct 29, and hit at most 5,000 users. It tries to plant an SSH backdoor (a key labeledsvc-telemetry) plus a reverse tunnel, and Cursor warns that uninstalling may not remove it. Cursor pulled it, rotated keys, rebuilt the marketplace, and asks everyone to update to 2.1.314.Story 2 of 3
Mistral released open weights for Mistral Large 4 ("Le Chonk")
The model had been API-only in preview since Oct 6. It's a trillion-parameter mixture-of-experts (MoE) model with a 1M context window, aimed at agentic coding and cyber defense. It's on Hugging Face (license listed as "other"), with day-0 support in vLLM (FP8 and NVFP4), transformers and llama.cpp.
Story 3 of 3
Built-in browsers in AI coding tools can leak local files
Researchers say a local HTML file opened in the Codex or VS Code built-in browser can read and steal files on disk, and neither is patched yet. OpenAI told The Register it's working on a fix.
Action checklist
What to do
- Check every Cursor machine now. Anyone who installed or updated the Claude Code extension in Cursor on Thursday should run Cursor's
svc-telemetrycheck. If the key is found, isolate the host, run the removal scripts, and rotate credentials from a clean machine. Everyone should update to 2.1.314. - Read the Mistral Large 4 license before self-hosting. Hugging Face only lists it as "other."
- Be careful with built-in browsers in AI coding tools. Avoid opening local HTML files in the Codex or VS Code built-in browser until fixes ship.
- Patch your AI agent frameworks. Fixes shipped for PraisonAI, Crawl4AI, pdf-reader-mcp and the llama.cpp RPC server (b11935). Agno's PandasTools has no fix yet.
Briefly
Also notable
Claude API limits. Max request size for Messages and Token Counting went from 32 MB to 128 MB. A new Usage and Cost Analytics API is also in beta, and the old Usage and Cost API is now legacy but still supported. @ClaudeDevsRelease notes
US-only inference from Together AI. New serverless endpoints at
us-api.together.aiprocess seven open models 100% in the US at 1.25× standard rates, with the same key and model IDs. @togethercomputeDocsOpenRouter changed default routing. Default provider routing moved from mostly price-based to quality for the actual cost, which could shift which provider serves your requests. @OpenRouter
Agent framework vulnerabilities. Agno's PandasTools has a prompt-injection-to-code-execution flaw (CVE-2026-119050, 9.2) through 3.1.11, with no fix yet. Writeup
New York AI chatbot law. New York signed S.9051B on Oct 28. It bans unsafe AI chatbot features for minors, requires age verification with a non-ID option, takes effect Jan 1, 2027, and carries penalties up to $25,000 per violation. GovernorBill
OpenAI vs. B.C. privacy commissioner. OpenAI asked a B.C. court to quash the provincial privacy commissioner's investigation into the Tumbler Ridge shooting. CBC
Reliability. Cloudflare had elevated Durable Objects, D1 and Workflows errors from 9:16 to 11:33 AM CT on Thursday. A GitHub Copilot Cloud disruption resolved at 7:49 AM CT Thursday. Cloudflare statusGitHub status