[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).
Summary
When one gem@version is installed in more than one gem home, vex (agent / manifest basis, with or without -g) verifies only the first copy the ruby crawler lists. That's always the gem env gemdir home (the Ruby install's gem dir, or GEM_HOME). RubyGems and Bundler load the copy from the first Gem.path entry instead. Gem.path lists the user gem dir (gem install --user-install) and every GEM_PATH entry before GEM_HOME. So when the loaded copy is unpatched and the gemdir copy is patched, vex writes not_affected for code that is actually vulnerable at runtime. In the reverse case (loaded copy patched, gemdir copy not), it refuses to attest a patched install.
apply / get / rollback are fine. They fan out to every copy (#222), so right after get -g both copies are patched. The bug appears as soon as the loaded copy is reinstalled: gem install --user-install <gem> again, gem pristine --user-install, or a reinstall into a GEM_PATH home.
Impact
A false OpenVEX not_affected statement (inline_mitigations_already_exist) for a package whose loaded code is unpatched. This is the "VEX attestation for a patch that isn't actually applied" class. It hits both vex -g and plain vex in a Bundler project that uses system gems (no BUNDLE_PATH), because bundle exec also loads the user-dir copy.
Repro (Linux; a local mock of the patch API serves one patch for colorize@0.8.1 that appends # SOCKET-PATCHED to lib/colorize.rb)
gem install colorize -v 0.8.1 --no-document # system gem home (gem env gemdir)
gem install --user-install colorize -v 0.8.1 --no-document # ~/.local/share/gem/ruby/X (first in Gem.path)
SP="socket-patch --api-url http://127.0.0.1:18999 --api-token fake --org org"
$SP get -g 11111111-2222-4333-8444-555555555555 --yes # patches BOTH copies (2 "Patched packages" lines)
gem install --user-install colorize -v 0.8.1 --no-document # user reinstalls -> the copy Ruby loads is pristine again
ruby -e 'require "colorize"; f=$LOADED_FEATURES.grep(/colorize.rb/)[0]; puts f, File.read(f)[/SOCKET-PATCHED/] ? "PATCHED" : "UNPATCHED"'
# /root/.local/share/gem/ruby/3.3.0/gems/colorize-0.8.1/lib/colorize.rb
# UNPATCHED
$SP vex -g --product pkg:gem/myapp@1.0.0 --output vex.json
# Using ruby gem paths at: /opt/rbenv/versions/3.3.6/lib/ruby/gems/3.3.0/gems
# Wrote OpenVEX document with 1 statement (exit 0)
grep -o '"status": "[a-z_]*"' vex.json
# "status": "not_affected"
The same thing happens with:
- an explicit
GEM_HOME=/tmp/gh GEM_PATH=/tmp/gp (Ruby loads /tmp/gp/..., vex -g checks /tmp/gh/...);
- a plain Bundler project (
gem "colorize", "0.8.1", no BUNDLE_PATH) with get <uuid> --mode agent and then vex. bundle exec loads the user-dir copy, and vex still says not_affected.
Reverse direction: when only the loaded user copy is patched, vex -g says omitting … (not_applied) and exits 1.
Expected vs actual
OS × version (probe run below, plus local Linux)
| OS |
Ruby / RubyGems / Bundler |
Gem.path[0] (loaded) |
vex -g after user-copy reinstall |
| Linux (sandbox) |
3.3.6 / 3.5.22 / 4.0.17 |
~/.local/share/gem/ruby/3.3.0 |
not_affected (also via GEM_PATH and plain vex) |
| ubuntu-latest |
2.7.8 / 3.1.6 / 2.4.22 |
~/.gem/ruby/2.7.0 |
not_affected |
| ubuntu-latest |
3.3.10 / 3.5.22 / 2.5.22 |
~/.local/share/gem/ruby/3.3.0 |
not_affected |
| ubuntu-latest |
3.4.9 / 3.6.9 / 2.6.9 |
~/.local/share/gem/ruby/3.4.0 |
not_affected |
| macos-latest (arm64) |
2.7.8 / 3.1.6 / 2.4.22 |
~/.gem/ruby/2.7.0 |
not_affected |
| macos-latest (arm64) |
3.3.10 / 3.5.22 / 2.5.22 |
~/.local/share/gem/ruby/3.3.0 |
not_affected |
| macos-latest (arm64) |
3.4.9 / 3.6.9 / 2.6.9 |
~/.local/share/gem/ruby/3.4.0 |
not_affected |
| windows-latest |
2.7 / 3.3 / 3.4 |
user gem dir |
not reached: global discovery itself is broken on Windows (#421) |
First bad version
This isn't a v5 regression. gem_env_gems_dirs already put gemdir first in v4.0.0 and v3.3.0 (checked in the source only).
Suspect code
crates/socket-patch-core/src/crawlers/ruby_crawler.rs:236-263 (gem_env_gems_dirs) pushes gem env gemdir first and then appends gem env gempath (which is already in Gem.path order and starts with the user dir / GEM_PATH). Using gempath order alone (it always contains gemdir) would match RubyGems.
crates/socket-patch-cli/src/commands/vex.rs:556 (collapse_to_first) verifies manifest-basis gem purls against one copy only. The hosted path in vex_consumed.rs already requires every gem copy to verify.
Probe run (macOS/Windows/Linux × Ruby 2.7/3.3/3.4): https://github.com/SocketDev/socket-patch/actions/runs/36816092864
[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).
Summary
When one
gem@versionis installed in more than one gem home,vex(agent / manifest basis, with or without-g) verifies only the first copy the ruby crawler lists. That's always thegem env gemdirhome (the Ruby install's gem dir, orGEM_HOME). RubyGems and Bundler load the copy from the firstGem.pathentry instead.Gem.pathlists the user gem dir (gem install --user-install) and everyGEM_PATHentry beforeGEM_HOME. So when the loaded copy is unpatched and the gemdir copy is patched,vexwritesnot_affectedfor code that is actually vulnerable at runtime. In the reverse case (loaded copy patched, gemdir copy not), it refuses to attest a patched install.apply/get/rollbackare fine. They fan out to every copy (#222), so right afterget -gboth copies are patched. The bug appears as soon as the loaded copy is reinstalled:gem install --user-install <gem>again,gem pristine --user-install, or a reinstall into aGEM_PATHhome.Impact
A false OpenVEX
not_affectedstatement (inline_mitigations_already_exist) for a package whose loaded code is unpatched. This is the "VEX attestation for a patch that isn't actually applied" class. It hits bothvex -gand plainvexin a Bundler project that uses system gems (noBUNDLE_PATH), becausebundle execalso loads the user-dir copy.Repro (Linux; a local mock of the patch API serves one patch for
colorize@0.8.1that appends# SOCKET-PATCHEDtolib/colorize.rb)The same thing happens with:
GEM_HOME=/tmp/gh GEM_PATH=/tmp/gp(Ruby loads/tmp/gp/...,vex -gchecks/tmp/gh/...);gem "colorize", "0.8.1", noBUNDLE_PATH) withget <uuid> --mode agentand thenvex.bundle execloads the user-dir copy, andvexstill saysnot_affected.Reverse direction: when only the loaded user copy is patched,
vex -gsaysomitting … (not_applied)and exits 1.Expected vs actual
vexonly attests what's applied to the code that runs.crates/socket-patch-cli/src/commands/vex_consumed.rsalready states the gem rule for hosted-basis purls: "bundler loads whicheverGem.pathhome it hits first: ALL must verify". fix(gem): crawl bundle-path roots in bundler precedence order, patch every coexisting copy, contain config-sourced roots #222 also says the single-representative consumers pick "the copy bundler actually loads".vexhashes only the crawler's first copy, and for gem homes that's always thegemdircopy, which isn't the oneGem.pathresolves first.OS × version (probe run below, plus local Linux)
Gem.path[0](loaded)vex -gafter user-copy reinstall~/.local/share/gem/ruby/3.3.0not_affected(also viaGEM_PATHand plainvex)~/.gem/ruby/2.7.0not_affected~/.local/share/gem/ruby/3.3.0not_affected~/.local/share/gem/ruby/3.4.0not_affected~/.gem/ruby/2.7.0not_affected~/.local/share/gem/ruby/3.3.0not_affected~/.local/share/gem/ruby/3.4.0not_affectedFirst bad version
This isn't a v5 regression.
gem_env_gems_dirsalready putgemdirfirst in v4.0.0 and v3.3.0 (checked in the source only).Suspect code
crates/socket-patch-core/src/crawlers/ruby_crawler.rs:236-263(gem_env_gems_dirs) pushesgem env gemdirfirst and then appendsgem env gempath(which is already inGem.pathorder and starts with the user dir /GEM_PATH). Usinggempathorder alone (it always containsgemdir) would match RubyGems.crates/socket-patch-cli/src/commands/vex.rs:556(collapse_to_first) verifies manifest-basis gem purls against one copy only. The hosted path invex_consumed.rsalready requires every gem copy to verify.Probe run (macOS/Windows/Linux × Ruby 2.7/3.3/3.4): https://github.com/SocketDev/socket-patch/actions/runs/36816092864