MOLO-STATIC-001: static-file jail escape via a symlinked intermediate directory
Scope: audit · README · AGENTS
In-house finding (not a public CVE), modeled on the general symlink-following path
traversal class (CWE-61, "UNIX Symbolic Link Following") that underlies many static
file server CVEs (for example nginx's historical alias-traversal advisories and
assorted CMS/static-server CVEs that rely on O_NOFOLLOW being assumed sufficient).
Where
src/runtime/static/disk_provider.cpp, DiskProvider::open:
int fd = openat(root_fd_, path.c_str(), O_RDONLY | O_NOFOLLOW | O_CLOEXEC);
O_NOFOLLOW only blocks the kernel from dereferencing a symlink at the final
component of path. If an earlier path segment is a symlink pointing outside the
jailed root, the kernel follows it transparently while resolving the rest of the path,
and openat succeeds. The realpath-based containment check a few lines below only
runs when this openat call fails and follow_symlinks_ (default false,
disk_provider.h) is true, so the strict, default configuration never reaches any
containment check at all for this shape of symlink.
Relevance
DiskProvider is MoloApp's jailed static-file backend, the piece responsible for
keeping /static/* requests inside the configured root. This defeats that jail with
a single symlink an attacker needs only the ability to place inside the served
directory tree (a common outcome of an unrelated file-upload or archive-extraction
feature elsewhere in an app, or simply a misconfigured deploy that leaves a
developer's symlink in place).
What the check does
Builds a scratch directory with www/escape -> outside/ (a symlink one level inside
the served root) and outside/secret.txt holding a marker string, starts
build/bin/runtime_interop_server on that root with follow_symlinks at its default
(false), and requests GET /static/escape/secret.txt. PASS means the jail held
(404); FAIL means the marker string came back.
Caveats
This does not test the follow_symlinks=true code path's own documented TOCTOU
window (that path already states its own limitation in a code comment); it tests the
supposedly-strict default, which the code comments do not flag as vulnerable to this
shape of symlink at all.
Run: make test in this directory, or make -C audit test from the repo root.