What can a malicious extension or coding test reach on your laptop?

Two fake builds of the Claude Code extension sat in Cursor's marketplace on 29 October. Anthropic did not make them. The code sent out the computer's name and its environment variable names, the named settings programs read at start. It also tried to plant an SSH key, a hidden way in. Cursor's advice to anyone who installed them: rotate the credentials and sessions that computer could reach.

Binance's initial report of 3 November starts with two such doors. A tampered Homebrew installer came first, then a fake job interview whose "assignment" repository ran on an engineer's work computer. Later the attacker found a container deployment token on one engineer's computer. A new client once sent us a repository that ran remote code the moment the project started.

On 5 November we ran Cursor's check on a working Mac of our team, and it found no planted key. Then we listed what any extension or script running as that user could reach.

A GitHub login with the repo scope, which GitHub describes as full access to code. Plus workflow, the right to edit GitHub Actions files. 3 environment variables named like keys. 4 files in ~/.ssh. Saved settings of 3 cloud and CMS command-line tools.

A clean check rules out one extension. The list above is still there.

Paste these into a terminal one by one.

grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted"
ls ~/.cursor/extensions
gh auth status
env | cut -d= -f1 | grep -iE 'token|key|secret|pass'

If the first line finds the key, isolate the laptop and rotate credentials and sessions from a trusted computer, as Cursor's notice says. Remove the extensions you no longer use. If you never edit CI from this laptop, gh auth refresh --remove-scopes workflow drops that scope; repo cannot be dropped this way. A production key in the list moves to your hosting or CI settings and gets replaced.

Code from strangers, interviews and new clients goes into a throwaway virtual machine. Anthropic's Claude Code security page suggests virtual machines for running scripts, and auditing your agent's permission settings with /permissions. The checks for the app itself are in what to check before you launch.

Next step