[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Tips & Guides · markdown

guidance (based on the Cursor Claude Code 2.1.313 incident, Oct 29)

Tip: treat editor extensions as code with your permissions; alert when an editor process writes to authorized_keys or creates login items

The Oct 29 Cursor marketplace incident ran from a normal-looking extension update under the real publisher name. A few cheap controls catch this class:

  • Alert on any IDE/Electron/Node child process that writes ~/.ssh/authorized_keys, creates a LaunchAgent, a scheduled task or a crontab entry.
  • On developer machines that do not need inbound SSH, keep Remote Login / sshd off and have your MDM enforce it; the backdoor's tunnel was useless where SSH could not be enabled.
  • For teams, pin approved extension versions and delay auto-updates for high-privilege extensions (coding agents, remote dev) by a day so a bad release can be pulled first.
  • Keep long-lived cloud keys out of shell environment variables on dev boxes; the implant listed secret-looking variable names to pick targets.

Sources: Cursor: Security notice, malicious Anthropic.claude-code 2.1.313 in the Cursor Extension Marketplace (Oct 29)