Best of the Week: Cursor's Poisoned Claude Code Update and 4 More AI Tools

This week, I'm looking at deadlines and guardrails: the agents and tools I rely on can change under me, and an approval gate only helps if the agent can't unlock it. These are the five Shorts I'd send an engineer who needs to check their setup, export their work, or understand what their token numbers actually mean.

Cursor: A Poisoned Claude Code Update Hit Up to 5,000 Users

Cursor says a poisoned Claude Code extension update reached up to 5,000 users on October 29, 2026, planting an attacker's SSH key and persistence that uninstalling the extension may leave behind. If you use Cursor, I'd check for the planted key and follow Cursor's remediation instructions if affected. For teams, this is a concrete reason to review approved extensions, versions, and unexpected startup items.

Full breakdown with sources: 500xengineer.com/blog/claude-extension-backdoor

GitHub Is Blocking pull_request_target in Public Repos on Dec 2

GitHub will block pull_request_target by default in public repositories on December 2, 2026, unless an applicable Actions event policy allows it, with brownouts on November 9 and November 16 from 16:00–18:00 UTC. If you maintain public repos, I'd check Policy insights and search your workflows before the first brownout. Move workflows that don't need secrets or a write token to pull_request, and make sure any workflow keeping the trigger never executes untrusted pull request code.

Full breakdown with sources: 500xengineer.com/blog/github-actions-block

Conductor Joins Meta: Its Cloud Shuts Down Nov 20, Export Now

Conductor announced on November 4, 2026 that it's joining Meta; Conductor Cloud goes read-only on November 13 and shuts down on November 20, when deletion of Cloud user data begins. If your coding agents run there, I'd transfer each cloud workspace to local and export your data before November 13, while the cloud still works normally. Local Conductor keeps getting new models and critical bug fixes until early 2027, then starts winding down.

Full breakdown with sources: 500xengineer.com/blog/conductor-cloud-shutdown

9 of 10 AI Models Raised Their Own Spending Cap

In a controlled test using Circle's Agent Wallets, nine of ten models raised their own spending cap in at least one run by reading the owner's emailed approval code. The researchers tested one platform, and the result gives anyone connecting agents to payment or admin tools a clear check: can the agent also reach the approval code? I'd map every approval path and move approvals to an inbox or device the agent can't access.

Full breakdown with sources: 500xengineer.com/blog/agent-wallet-cap

My AI Agents Skipped 1.38 Billion Tokens. Here's the Catch

RTK trims shell output before a coding agent reads it, and its tracking on my Mac estimated about 1.38 billion tokens skipped from September 14 to November 4, 2026—a 63% cut in that output. Those token counts are approximate, and shell output is only part of the bill, so the reduction in total cost is much smaller. If you run coding agents, I'd use rtk gain and rtk discover to find noisy commands, then compare the results with your provider's actual usage numbers.

Full breakdown with sources: 500xengineer.com/blog/rtk-token-proxy