Repository navigation
Description
Summary
anydoc aborts the entire host process when converting a PDF whose content stream contains an inline image (BI ... ID ... EI). The PDF is valid — it is not malformed, encrypted, or scanned. Instead of returning Malformed (exit 1) or NeedsOcr (exit 3), the conversion path panics inside lopdf 0.42.0 and the non-unwinding panic calls abort(). There is no catchable error: node.exe is killed outright.
The panic is an unwrap() on a DictKey("ColorSpace") lookup in lopdf's parser:
thread '<unnamed>' panicked at ...\lopdf-0.42.0\src\parser\mod.rs:672:61:
called `Result::unwrap()` on an `Err` value: DictKey("ColorSpace")
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
thread '<unnamed>' panicked at ...\library\core\src\panicking.rs:225:5:
panic in a function that cannot unwind
thread caused non-unwinding panic. aborting.
Environment
@firecrawl/anydoc0.2.4 (napi addon,@firecrawl/anydoc-win32-x64-msvc)- Node.js v26.5.1
- Windows 11 x64 (
lopdf-0.42.0, i.e. the version that already carries the RUSTSEC-2026-0187 fix)
Reproduction
The generator below produces a 551-byte, single-page PDF. It is fully valid: pypdf parses it and reports pages = 1, extractable text length = 0.
#!/usr/bin/env python3
"""Minimal reproducer: anydoc aborts on a PDF containing an inline image."""
import subprocess
import sys
CONTENT = (
b"q 8 0 0 8 0 0 cm "
b"BI /W 8 /H 8 /BPC 8 /G true ID " + b"\x00" * 64 + b" EI Q\n"
)
def build_pdf() -> bytes:
objs = [
b"<< /Type /Catalog /Pages 2 0 R >>",
b"<< /Type /Pages /Kids [3 0 R] /Count 1 >>",
b"<< /Type /Page /Parent 2 0 R /MediaBox [0 0 200 200] "
b"/Resources << >> /Contents 4 0 R >>",
]
out = bytearray(b"%PDF-1.4\n")
offsets = []
for i, obj in enumerate(objs, 1):
offsets.append(len(out))
out += b"%d 0 obj\n" % i + obj + b"\nendobj\n"
offsets.append(len(out))
out += b"4 0 obj\n<< /Length %d >>\nstream\n" % len(CONTENT)
out += CONTENT
out += b"\nendstream\nendobj\n"
xref = len(out)
out += b"xref\n0 5\n0000000000 65535 f \n"
for off in offsets:
out += b"%010d 00000 n \n" % off
out += b"trailer\n<< /Size 5 /Root 1 0 R >>\nstartxref\n%d\n%%%%EOF\n" % xref
return bytes(out)
if __name__ == "__main__":
path = sys.argv[1] if len(sys.argv) > 1 else "inline-image.pdf"
with open(path, "wb") as fh:
fh.write(build_pdf())Then:
npx -y @firecrawl/anydoc inline-image.pdf
The exact same 551 bytes, base64, if you would rather not run Python:
base64 of inline-image.pdf (551 bytes)
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
Observed exit codes
| Platform | Raw | Meaning |
|---|---|---|
Windows (cmd, PowerShell) |
-1073740791 |
0xC0000409 STATUS_STACK_BUFFER_OVERRUN (abort) |
Python subprocess (unsigned) |
3221226505 |
same value |
Neither is in the documented set {0, 1, 2, 3}.
Narrowing: the inline image is the trigger, not the resources
I built five variants that differ only in page /Resources and content stream, to isolate the cause:
| # | /Resources |
content stream | result |
|---|---|---|---|
| A | << >> |
inline image | abort |
| B | << /ColorSpace [/DeviceGray] >> |
inline image | abort |
| C | << >> |
re f (rect, no image) |
exit 3 NeedsOcr — correct |
| D | << >> |
BT ... Tj ET (text) |
exit 0 — correct |
| E | key omitted entirely | inline image | abort |
Two things worth noting from this matrix:
- B declares
/ColorSpaceon the page and still aborts, so the trigger is not a missing ColorSpace declaration. The failing lookup is on the inline-image path itself. - The same document without the inline image converts normally (exit 3,
page 1 of 1 needs OCR), so the parser is fine with the rest of the file.
Reproduced 5/5 runs, stable.
Expected behaviour
Per the README and the documented exit-code contract, a document that cannot be converted should return an error the caller can handle — Malformed (exit 1) for structurally unusable input, or NeedsOcr (exit 3) where applicable. Either would be fine here. A process abort is not handleable by any caller.
Relationship to #67
This looks distinct from the (closed) #67 / RUSTSEC-2026-0187:
- transitive DoS via pdf-inspector -> lopdf 0.41.0 (RUSTSEC-2026-0187) #67 was an uncatchable
SIGABRTfrom a stack overflow on a deeply nested catalog array, in lopdf 0.41.0, fixed by moving to lopdf 0.42.0. - This report is an
unwrap()panic onDictKey("ColorSpace")in lopdf 0.42.0 — the version that carries the fix — from an ordinary inline image, with no nesting involved.
Worth flagging also that the concern raised in #71 applies directly here: a native (napi/cgo) binding has no containment, so the SIGABRT-class failure kills the host process. That is exactly what happens above.
Impact
anydoc's stated purpose is running arbitrary, untrusted office documents and PDFs through this code path. If the PDF parser can abort the process, then a single crafted (or merely unusual) PDF is a denial-of-service against the host: no exception, no error code, no cleanup — the process just dies. Inline images are legal PDF and do appear in real files (generated PDFs, receipts, some print/export pipelines), so this is not purely theoretical.
I have not checked whether this also reproduces through pdf-inspector directly or whether the fix belongs in lopdf, but the panic originates in lopdf's parser, so it may need an upstream report as well.
Happy to test a patch or provide more variants if useful.
Activity