VulnWatch VulnWatch
← Back to dashboard
Medium github · GHSA-xmfg-f9cm-w86q

NLTK: Symlink escape in chat80.sql_query bypasses the pathsec data-root sandbox to read out-of-root SQLite databases

Published Nov 16, 2026 CVSS 6.5

Summary

nltk.sem.chat80.sql_query() opens a SQLite database by handing the result of nltk.data.find(dbname) straight to sqlite3.connect() without the path-containment check that the NLTK data sandbox (nltk.pathsec) is meant to enforce. nltk.data.find() applies only lexical checks to the resource name and returns a FileSystemPathPointer whose path is normalized with os.path.abspath but never resolved for symlinks. As a result, an in-root symbolic link that points at a SQLite file outside every registered NLTK data root is followed by SQLite, and the caller receives query results from the out-of-sandbox database. An attacker who can place a file inside a directory the application trusts as an NLTK data root can read SQLite databases that the pathsec sandbox is supposed to keep out of reach.

Root Cause

The security invariant is that file access performed through NLTK data resolution must stay within the configured data roots; the sandbox enforces this in nltk.pathsec via validate_path and pathsec.open, which resolve the real path (Path(...).resolve()) and reject targets outside the allowed roots. sql_query() never reaches that enforcement point. It calls nltk.data.find(dbname), whose guards operate purely on the resource-name string, and FileSystemPathPointer.__init__ performs only os.path.abspath plus an existence check, which does not collapse symlinks. The pointer is then passed to sqlite3.connect(str(path)), so the containment decision that FileSystemPathPointer.open() would delegate to pathsec.open is bypassed entirely. The sibling functions in the same module that create or write databases (cities2table, val_dump, val_load) obtain their protection from a validation call in later development, but the read sink sql_query performs no equivalent check in any published release.

nltk/sem/chat80.py  -> sql_query(): path = nltk.data.find(dbname); sqlite3.connect(str(path))
nltk/data.py        -> find(): lexical name guards only; returns FileSystemPathPointer
nltk/data.py        -> FileSystemPathPointer.__init__(): os.path.abspath + os.path.exists (no symlink resolution)
nltk/data.py        -> FileSystemPathPointer.open() -> nltk.pathsec.open (the containment guard that sql_query never calls)

Impact

An attacker who can write into a directory the application registers as an NLTK data root (a shared or writable corpus directory, or a downloaded corpus) can plant a symbolic link at the expected resource location that points at a SQLite database elsewhere on the filesystem. When the application invokes sql_query() (directly or via sql_demo()) for that resource, SQLite follows the link and the query executes against the out-of-sandbox database, returning its rows to the caller. This escapes the pathsec confidentiality boundary. The read primitive is bounded to SQLite-format files that are readable by the process user: a non-SQLite target raises sqlite3.DatabaseError rather than leaking content. A caller-supplied write query would be redirected onto the out-of-sandbox target and SQLite may create sidecar files (-wal, -journal) next to it, which is a plausible integrity side effect rather than a demonstrated arbitrary-write primitive. Exploitation requires the same local write capability into a trusted data root as the related symlink-read sandbox-escape issues already recorded for this project.

Proof of Concept

Requires the published nltk 3.10.3 wheel and a writable directory registered as an NLTK data root. No network access is needed.

  1. Create a throwaway data root with the expected corpus subdirectory.
  2. Create a SQLite database outside that root and populate it.
  3. Plant an in-root symbolic link at corpora/city_database/city.db pointing at the outside database.
  4. Register the data root and call sql_query() for the in-root resource name.
import os, tempfile, sqlite3, nltk, nltk.data
from nltk.sem import chat80

work = tempfile.mkdtemp()
root = os.path.join(work, "nltk_data")
corp = os.path.join(root, "corpora", "city_database")
os.makedirs(corp)

secret = os.path.join(work, "outside", "secret.db")
os.makedirs(os.path.dirname(secret))
con = sqlite3.connect(secret)
con.execute("CREATE TABLE t (x text)")
con.execute("INSERT INTO t VALUES ('OUT-OF-SANDBOX-SECRET')")
con.commit(); con.close()

os.symlink(secret, os.path.join(corp, "city.db"))
nltk.data.path.insert(0, root)

ptr = nltk.data.find("corpora/city_database/city.db")
try:
    ptr.open().read()            # pathsec route: blocked
except Exception as e:
    print("open() blocked:", type(e).__name__)

cur = chat80.sql_query("corpora/city_database/city.db", "SELECT * FROM t")
print("sql_query rows:", cur.fetchall())   # sink route: follows the symlink
open() blocked: PermissionError
sql_query rows: [('OUT-OF-SANDBOX-SECRET',)]

Attack Chain

  1. Exposure: The application exposes sql_query() (or sql_demo(), which calls it) and resolves the target database through nltk.data.find, using at least one directory the attacker can write to as a data root.
  2. Control: The attacker controls the on-disk object at the in-root resource path, replacing it with a symbolic link whose target is chosen freely. The resource-name argument stays a clean, in-root string.
  3. Path: nltk.data.find applies its lexical name guards (_UNSAFE_NO_PROTOCOL_RE, encoded and normalized-form checks), which inspect only the name string and pass, then returns a FileSystemPathPointer whose __init__ applies os.path.abspath and an existence check without resolving the symlink.
  4. Guard: The containment guard (FileSystemPathPointer.open() delegating to nltk.pathsec.open, which resolves the real path and rejects out-of-root targets) is never invoked, because sql_query() passes the raw pointer string to sqlite3.connect() instead of opening through the pointer.
  5. Primitive: sqlite3.connect(str(path)) opens the symlink target; the kernel follows the link out of every data root and the cursor executes the query against the outside database.
  6. Result: Rows from a SQLite database outside the sandbox are returned to the caller, escaping the pathsec confidentiality boundary for SQLite-format files readable by the process user.

Bypass Evidence

The relevant guard is the pathsec containment enforced by nltk.pathsec.open/validate_path, reached through FileSystemPathPointer.open(). On an identical pointer produced from the planted symlink, ptr.open() raises PermissionError with an "Unauthorized path" security violation, confirming the guard is active and correct for the sandbox route. sql_query() does not call .open() and does not call validate_path; it passes the unresolved pointer string directly to sqlite3.connect(), so the guard never runs on this path and the read succeeds. Attempts to invalidate the finding were made and failed: nltk.data.find was checked for a hidden containment or real-path resolution and has none (its guards are name-string only, and FileSystemPathPointer.__init__ uses abspath, not realpath); SQLite was confirmed to follow the link and read the outside file rather than opening the link node; and the possibility that an existing published fix already covers this sink was refuted by inspecting the published artifacts, where no containment check is present on sql_query. The behavior was reproduced against the exact published 3.10.3 wheel (verified by its published SHA-256 digest).

Affected Versions

  • Ecosystem: PyPI
  • Package: nltk
  • `Confirmed affected range: >= 3.9.4,

Affected AI Products

cursor
Get the weekly digest. Every Monday: top AI security stories of the week. Free.