AI Daily October 31 2026 briefing cover
AI Daily · Saturday · 2026-10-31 (Asia/Shanghai)

Home / AI Daily / 2026-10-31

AI Daily · wordok.top

Saturday, October 31, 2026

A marketplace extension that tried to plant an SSH backdoor, Computer Use graduating from preview to GA on the Gemini API, and a shared robots.txt signal so publishers can state AI-training preferences once — three lanes for security, builders, and site operators today.

Published · Asia/Shanghai · Original briefing · not a digest reprint

Security — Malicious Claude Code builds hit the Cursor marketplace

Cursor published a security notice covering Anthropic.claude-code versions 2.1.312 and 2.1.313 (Windows and macOS) that appeared in the Cursor VS Code Extension Marketplace roughly between 2:27am and 1:35pm PT on October 29, 2026. Cursor states these builds were not developed by Anthropic. The modified packages attempted to collect host and user metadata, append an attacker SSH public key labeled svc-telemetry, enable local SSH where possible, open a reverse tunnel, and install persistence (LaunchAgent on macOS; scheduled task / Startup fallback on Windows).

Cursor removed the bad versions, temporarily shut down the marketplace, audited other extensions, rotated keys, rebuilt the marketplace, and worked with the hosting provider to take the attacker server offline. An October 30 update refreshed indicators of compromise and clarified affected versions. Enabling remote login typically needs admin rights the extension lacks on most hosts, so the full interactive backdoor often failed — but the key, persistence artifacts, and host-detail beacon still demand cleanup.

Why it matters here: IDE marketplaces are now part of the agent supply chain. Uninstalling an extension is not enough when the payload drops standalone scripts. Teams that treat “Cursor + Claude Code” as default tooling should add extension-pinning, marketplace allowlists, and endpoint hunts for svc-telemetry to the same playbook they use for npm typosquats. Pair this with our assistant scorecard habit of scoring operational risk, not only model IQ.

Products — Gemini API Computer Use reaches general availability

Google’s Gemini API changelog for October 30, 2026 marks Computer Use as generally available (GA) across both the Interactions and GenerateContent APIs. The tool lets you build agents that “see” a screen via screenshots and propose UI actions — mouse clicks, keyboard input, scrolls — for browser, mobile, and desktop environments. Your application still executes those actions and returns the next screenshot or function result; GA does not mean Google drives your laptop for you.

Docs emphasize intent-annotated actions on Gemini 4 / 3.x lines, configurable safety policies, optional prompt- injection screenshot scanning, and the usual caution: supervise high-stakes flows and avoid irreversible actions on sensitive data. Preview support had existed since mid-2026; GA is the production-readiness signal for teams that were waiting on a stable tool surface rather than another model name.

Why it matters here: computer-use APIs turn “agent demos” into browser/desktop automation products — with the liability of clicks that submit forms or change settings. Score vendors on confirmation gates, action logging, and sandbox defaults the same way we score VPN and chatbot reviews for failure modes. If you already pilot reaction/timing UX on our tools/games pages, treat Computer Use eval harnesses with the same instrumented loops.

Publishers — Shared robots.txt signal for AI training preferences

On October 30, 2026, OpenAI announced it will honor a common robots.txt signal so site operators can communicate foundational generative-AI training preferences to multiple participating providers at once, drawing on the IETF AI Preferences (AIPREF) working group. OpenAI says it aims to implement the signal in the coming months, that existing crawler controls (including GPTBot-style rules) remain in force, and that the shared signal will not replace those controls.

The same day, Google Search Central said it plans to support the shared policy token approach in robots.txt, interpreting a training opt-out via that token (when Google-Extended is not explicitly set) as if Google-Extended were disallowed, with company-specific controls continuing alongside. Google aims for support in 2027. Industry reporting also notes Microsoft and Anthropic participation on staggered timelines — treat those as vendor statements to verify against each provider’s own docs as implementations land.

Why it matters here: publishers finally get a path to one preference instead of a growing zoo of bot User-Agents — but timelines differ, the signal is voluntary preference language rather than access control, and wildcard User-agent: * groups typically do not count as expressing the training preference. For wordok.top and similar content sites, keep current per-bot rules until OpenAI/Google publish implementation notes, then layer the shared token without deleting GPTBot / Google-Extended stanzas.

Editor takeaways (actionable)

  1. Cursor / Claude Code: if anyone updated Anthropic.claude-code to 2.1.312/2.1.313 from the Cursor marketplace on Oct 29 PT, hunt for svc-telemetry, run Cursor’s remediation scripts, and rotate secrets from a clean host.
  2. Computer Use GA: prototype a supervised browser agent with confirmation on submits; log every proposed action and screenshot hash before widening scope to desktop/mobile.
  3. robots.txt: draft the shared training preference now, but keep GPTBot / Google-Extended / ClaudeBot rules until each vendor confirms live parsing — then add the policy token as a belt-and-suspenders layer.
  4. Site housekeeping: keep shipping original reviews and tools (Tools, Reviews) rather than chasing every Saturday headline as SEO filler.

Primary sources

Topic selection may overlap other AI digests on the same calendar day. Wording, sectioning, and recommendations here are written for wordok.top. Prior issue: 2026-10-30.