Cursor's Marketplace Served a Backdoored Claude Code Extension

Act before next issue

On our radar

• Muse Wonder Preview at $4/$20 per 1M tokens

  Meta released Muse Wonder on 12 October as a preview on the Meta Model API (muse-wonder-1.0), with a 1,048,576-token context. The stronger Muse Wonder Full is only in Muse Code's Power Usage and Maximum Usage plans. Every benchmark result is Meta-reported, and we have not tested either.

• LiteLLM GHSA-g5ff-637f-6q2m: upgrade to 1.95.0

  LiteLLM's 1 October advisory rates an arbitrary file read High (CVSS 8.1, no CVE): an authenticated internal_user_viewer can pull environment variables and API keys off the proxy. Versions before 1.95.0 are affected and there is no configuration workaround.

• Vercel Eve 0.78.0 closes a privilege escalation

  Release notes for eve@0.78.0, dated 13 October, say approved calls now always run as the requester. Before, the approver's identity and credentials could be used, letting a requester reach access they did not have.

• In review, not yet published

  None of these drafts has passed our source review, so we report no figures from them yet: GitHub Copilot's model deprecations, the Gemini API's storage default for new paid projects, the Cloudflare Workers AI GLM-5.3 exposure, planted AGENTS.md and CLAUDE.md files in Downloads folders, Claude Code 2.1.311, GPT-6.1 Terra, LiteLLM's attack-chain root cause and its Private Registry, Kimi K3.1 Pro, OpenAI's offline frontier training, Cohere's Qwen3.8-Flash-Next findings, unified Claude Enterprise, Codex CLI 0.164 to 0.166, the x402 SDK spending cap, JetBrains ThinkRail, Mistral Large 4 weights and pricing, API credits for ChatGPT Pro and Business, OpenAI's Evals and Agent Builder shutdowns, Gemini 4 Carbon, Claude Code's background status line cost, GitHub Copilot Fusion, Copilot in GitHub Enterprise, Nemotron 3.5 Transcribe, Cursor's Agent SDK, Artificial Analysis Intelligence Index v5, ElevenLabs API deprecations, leaked keys on Gitee, fake desktop app repos, GitHub's npm package firewall, Anthropic's exploit-chain disclosure, Xint's compaction finding and Liquid AI LFM3. Muse Spark Open and a DeepSeek peak-window correction passed review but are not yet live. Our GPT-6 Astra figures correction was updated on October 31 and is live; Astra stays conditional.

• Claude Code extension scope claims

  Some early reports said Open VSX or Microsoft Marketplace builds of the extension were also affected. Cursor and OpenSourceMalware say only Cursor's marketplace served them, and Anthropic has made no statement. Held until Anthropic or the registries confirm either way.

• Deprecations and allowances still unconfirmed

  OpenAI's Help Center page on OpenAI models in Cursor could not be read this week, so the wind-down still stands as proposed with a 12 November transition period and no termination date. A blog claim that OpenAI halved the Codex allowance on its $200 tier has no OpenAI page behind it yet.

• Security items with no AI tool named

  Ransomware affiliates using an autonomous red team framework, two Wasmtime sandbox escapes, Apple's macOS guidance on the compromised eMudhra certificates and hosted providers for abliterated DeepSeek models. None names a model, harness or tool a reader runs. Held until one does.

• Launches read only as a title or feed summary

  OpenAI's updated Model Spec and its post on detecting misalignment, Google's developer platform for Gemini and Search, Opera WebAI, Mistral's cybersecurity variant of Large 4, Claude Enterprise smart reports, the Gemini API CLI beta, Block's Dessa speech model and the Hurston-1 TTS model. Held until a primary page shows a model, price, release or behaviour change.

• Policy, litigation and research held at headline level

  Swedish government offices banning ChatGPT and Claude, Google's Gmail and Photos ad policy, the Stanford authors' claim against Anthropic, a lawsuit over GPT-4o advice, a Wikimedia warning on an AI copyright ruling, Deepgram buying Daily (maker of Pipecat) and a claim that five Qwen3.8-27B models match Fable 5. Nothing read changes a product yet.

• Open verdict watches

  Moves of GPT-4.1 and GPT-4o to caution are staged from OpenAI's removal dates and land only when a carrying post publishes. Mistral Large 4 stays conditional after its weights shipped, and LiteLLM and the OpenAI Decisions API stay conditional with refreshed summaries staged.