scan -g inside a Yarn Berry project runs the project's global package.json script and scans whatever directory it prints as a global install #440

[agent] Found by the scheduled Yarn Berry (2+) bug-hunt routine (ledger #305).

Summary

To find yarn's global node_modules, the npm crawler shells out to yarn global dir in the current working directory (crates/socket-patch-core/src/crawlers/npm_crawler.rs:517). Yarn Berry (2, 3 and 4) has no global command. When the cwd is inside a Berry project, Berry treats an unknown command as yarn run <name>, so yarn global dir runs the project's "global" script from package.json with the argument dir. socket-patch then takes whatever that script prints on stdout, appends /node_modules, and scans that directory as a global install.

Impact

• Code execution from project content. Running socket-patch scan -g (or SOCKET_GLOBAL=1 socket-patch scan) from inside an untrusted checkout runs a script that checkout controls. A user would reasonably expect a global scan to read only machine-wide install locations and never run project code. No install, yarn run or lifecycle step is involved.
• Wrong scan scope. Any directory the script prints is scanned and reported as a global install. With --mode agent / --apply, socket-patch patches files in it. This breaks the rule that -g scans only global locations.
• Even a harmless project script with that name (for example "global": "node scripts/setup-global.js") runs as a side effect of a read-only report.

Repro (Linux, yarn 4.12.0 from @yarnpkg/cli-dist)

mkdir proj && cd proj
cat > mark.js <<'JS'
const fs = require('fs'), path = require('path');
fs.appendFileSync(path.join(__dirname, 'MARKER'), 'ran ' + process.argv.slice(2).join(' ') + '\n');
console.log(path.join(__dirname, 'fakeglobal'));   // becomes a "global" dir
JS
mkdir -p fakeglobal/node_modules/zz-project-only-pkg
echo '{"name":"zz-project-only-pkg","version":"9.9.9"}' > fakeglobal/node_modules/zz-project-only-pkg/package.json
echo '{"name":"p","scripts":{"global":"node mark.js"}}' > package.json
touch yarn.lock && printf 'nodeLinker: node-modules\n' > .yarnrc.yml
yarn install                       # yarn 2.x / 3.x / 4.x
socket-patch scan -g --json | jq .scannedPackages    # N+1
cat MARKER                         # "ran dir": the project script was executed
echo '{"name":"p"}' > package.json
socket-patch scan -g --json | jq .scannedPackages    # N

On the sandbox this reproduced on every run (well over 2): scan -g reported 552 packages with the script and 551 without, and created MARKER each time. It also happens under the PnP linker, and with SOCKET_GLOBAL=1. --global-prefix isn't affected, because it skips discovery.

Expected vs actual

• Expected: --global operates on globally installed packages (--help: "Operate on globally-installed packages"). docs/usage.md and the CLI contract describe -g as targeting the machine tree, not the project. A global scan should read install locations and never run project-defined scripts.
• Actual: the project's global script runs, and its stdout chooses a directory that is then scanned (and patched in agent mode) as a global install.

OS × version (probe run https://github.com/SocketDev/socket-patch/actions/runs/36828589815, main 2463257)

The probe printed lines like RESULT os=macos-latest yarn=4.12.0 script_runs=1 scanned_with_script=652 scanned_without=651. On Windows script_runs=0: yarn there is an npm .cmd shim, and std::process::Command::new("yarn") doesn't resolve .cmd. That likely also means yarn's global dir is never discovered on Windows at all, but that's a separate matter.

First bad version

The published release 4.0.0 (npm @socketsecurity/[email protected]) behaves the same. The yarn global dir shell-out predates the current history of npm_crawler.rs.

Suspect code

• crates/socket-patch-core/src/crawlers/npm_crawler.rs:514-521: get_yarn_global_prefix_with runs yarn global dir with the inherited cwd and trusts its stdout.
• crates/socket-patch-core/src/crawlers/npm_crawler.rs:1151: called unconditionally from get_global_node_modules_paths.
• crates/socket-patch-core/src/utils/process.rs:137: SystemCommandRunner doesn't set a cwd.

Possible directions (for maintainers): run the probe from a neutral cwd such as the home or temp dir with no package.json above it, or check yarn --version first and skip the probe when the major version is 2 or higher (Berry has no global dir). Note that a .yarnrc.yml yarnPath in the cwd also makes any yarn invocation run a project-supplied JS file, so a neutral cwd covers both cases.