Pillow: Memory disclosure and segfault in libtiff encoder via crafted TIFF tags

============================================================================
Memory disclosure and segfault in libtiff encoder via crafted TIFF tags
Location: src/encode.c:1014  in PyImaging_LibTiffEncoderNew()
Commit:   a707c6b21c313880d0771e54965d142dd42358b6
============================================================================

A crafted TIFF file of a few hundred bytes makes Pillow copy an attacker-chosen amount of its own process memory into the TIFF it writes back out, or crash the process outright. The trigger is an ordinary Image.open() followed by a bare im.save(fp, "TIFF"). The only precondition is that the application re-saves images as TIFF. When the uploaded file is compressed, Pillow routes that save through libtiff automatically, with no options from the application. The attached reproducer does this with a 356-byte LZW TIFF: one crafted tag comes back holding 65536 bytes taken from _imaging.so's read-only data, and a second crafted tag makes the worker die with SIGSEGV on every run.

The bug is in PyImaging_LibTiffEncoderNew (src/encode.c). It decides what arguments to pass to the variadic TIFFSetField() from the shape of the Python value (scalar, tuple or bytes). It assumes its ImagingLibTiffMergeFieldInfo() call defined how libtiff will read them. But TIFFMergeFieldInfo() never overrides libtiff's built-in field table, so libtiff reads the arguments according to its own definition of the tag:

• For a counted array tag such as DNGPrivateData (50740), a count-1 integer in the source file is taken as the byte count. The data pointer is then read from an argument slot Pillow never wrote, and that many bytes are copied into the output.
• For a fixed-size array tag such as ActiveArea (50829), the integer itself is used as the source address.

Where that unwritten slot points depends on the compiler. In the tested clang build it points into _imaging.so, and a larger count returned thousands of pointers into libpython, libc and libtiff, which defeats address-space randomisation for anyone chaining another bug. When the application preserves metadata with tiffinfo=im.tag_v2, a 64-bit entry turns ActiveArea into a 16-byte read from an address the attacker chooses. All observed memory errors are reads; no out-of-bounds write was found.

Other branches of the same loop share the defect and are reachable from crafted TIFFs or from EXIF in JPEG/PNG/WebP re-saved as compressed TIFF. These are the bytes, RATIONAL-tuple and typed-tuple branches, and a BYTE-typed ColorMap. The file-reachable forwarding was introduced in a3d45e9cefd0 ("Added custom int and float TIFF tags") and first released in 5.4.0.

Details

PyImaging_LibTiffEncoderNew (src/encode.c:777-1060) chooses the argument list it hands to the variadic TIFFSetField() from the Python value of each tag — scalar vs. tuple vs. bytes, and a type that is either autodetected or copied from the source file's IFD — and assumes that its own ImagingLibTiffMergeFieldInfo() registration defines how libtiff will read the va_list. It does not: libtiff keeps its built-in definition for every tag it already knows, and core tags are never registered at all. libtiff therefore consumes the varargs as (count, pointer), (pointer) or three pointers according to its field table, whatever Pillow actually pushed. The encoder never consults that table (TIFFFindField / TIFFFieldPassCount / TIFFFieldWriteCount / TIFFFieldDataType) before the call.

The most severe instance, and the one the attached reproducer exercises: when a TIFF is re-saved through the libtiff writer, every tag of the source IFD that Pillow does not recognise is forwarded to libtiff with TIFFSetField(). A count-1 integer entry is passed as a single scalar vararg. If libtiff's own field table defines that tag number as an array field, libtiff consumes the varargs as (count, pointer) or (pointer) instead, so the attacker's integer becomes either the number of bytes to copy from a never-written vararg slot, or the source address itself. The same mismatch is reachable through the typed tuple branches (tuple length consumed as a pointer for fixed-count built-in tags such as CFARepeatPatternDim, even when the source file is well-formed), the bytes, TIFF_RATIONAL-tuple and core-tag branches of the same loop; those are listed under Other branches with the same defect below and need to be covered by the same fix.

A 181-byte TIFF (16×16 RGB, one LZW strip, plus one extra IFD entry 50740 / LONG / count 1 / value 4096 — DNGPrivateData) opened with Image.open() and saved with a bare im.save(buf, "TIFF") produces a 4296-byte output whose tag 50740 holds 4096 bytes that occur nowhere in the input; scanning /proc/self/mem finds them in the read-only data segment of _imaging.so. Changing the extra entry to 50829 / LONG / 1 / 64 (ActiveArea) makes the same two calls segfault in memcpy(dst, src=0x40, n=16) under ImagingLibTiffSetField ← PyImaging_LibTiffEncoderNew. An ASan build of the extensions reports READ of size 4096 in memcpy called from src/encode.c:1014 for the first input.

Code path

_save copies all tags of the opened image into the libtiff attribute list. For a tag that is neither in LIBTIFF_CORE nor in TiffTags.TAGS_V2, no type is recorded, so the value goes through untyped:

# src/PIL/TiffImagePlugin.py:1782
supplied_tags = {**legacy_ifd, **getattr(im, "tag_v2", {})}
…
# :1959-1982
for tag, value in itertools.chain(ifd.items(), supplied_tags.items()):
    if tag not in TiffTags.LIBTIFF_CORE:
        …
        elif isinstance(value, (int, float, str, bytes)) or …:
            type = TiffTags.lookup(tag).type      # None for 50740, 50829, …
            if type:
                types[tag] = type
    if tag not in atts and tag not in blocklist:
        …
            atts[tag] = value                     # plain int 4096

compression defaults to the source image's im.info["compression"] (TiffImagePlugin.py:1731-1746), so a compressed source selects the libtiff writer without the application asking for it.

In the encoder the untyped int is autodetected as TIFF_LONG (src/encode.c:805-815), the tag is "registered" as a scalar custom field, and then set with one 32-bit vararg:

// src/encode.c:844      (is_var_length = 0 → writecount 1, passcount 0)
if (ImagingLibTiffMergeFieldInfo(&encoder->state, type, key_int, is_var_length)) {
    … continue;
}
…
// src/encode.c:1013-1016
} else if (type == TIFF_LONG || type == TIFF_IFD) {
    status = ImagingLibTiffSetField(
        &encoder->state, (ttag_t)key_int, (UINT32)PyLong_AsLong(value)
    );

The registration does not do what the caller assumes. TIFFMergeFieldInfo() only adds definitions for tags libtiff does not already know and returns 0 (success) either way, so for tags in libtiff's built-in table the built-in definition stays in force. Queried against the linked libtiff 4.5.0 after exactly Pillow's merge call:

50740 DNGPrivateData      type=BYTE   passcount=1 writecount=-1   (unchanged)
50829 ActiveArea          type=LONG   passcount=0 writecount=4    (unchanged)
65000 CustomField         type=LONG   passcount=0 writecount=1    (Pillow's definition, tag was unknown)

libtiff's setter for FIELD_CUSTOM tags then interprets the va_list according to its definition: for passcount=1 it reads va_arg(ap, int) as the element count — the attacker's 4096 — allocates count * size bytes and memcpys that many bytes from va_arg(ap, void *), a slot PyImaging_LibTiffEncoderNew never wrote. On x86-64 that slot is register rcx; in this build it holds a pointer into _imaging.so's .rodata, so the copy succeeds deterministically (a value of 2 000 000 also succeeds and yields a 2 000 200-byte output). TIFFWriteDirectory writes the copied bytes into the output file as the tag's value. For fixed-count array fields (passcount=0, writecount=4 for ActiveArea) there is no count argument, so the attacker's scalar is consumed as the void * source and 16 bytes are read from address 64 → SIGSEGV.

Tags in libtiff 4.5's built-in table that Pillow forwards untyped and that take the (count, pointer) form include DNGPrivateData 50740, CFAPattern 33422, CFAPlaneColor 50710, LinearizationTable 50712, BlackLevel 50714, WhiteLevel 50717, ColorMatrix1/2 50721/50722, AsShotNeutral 50728, OriginalRawFileName/Data 50827/50828, AsShotICCProfile 50831, CurrentICCProfile 50833. Fixed-count ones (pointer only → crash) include ActiveArea 50829, DNGVersion/DNGBackwardVersion 50706/50707, DefaultScale/DefaultCropOrigin/DefaultCropSize 50718–50720, AsShotWhiteXY 50729, CFARepeatPatternDim 33421, BlackLevelRepeatDim 50713, ImageLayer 34732, RawDataUniqueID 50781.

When the application saves with tiffinfo=im.tag_v2 (the documented way to preserve metadata), the file-declared type is honoured, and a LONG8 entry reaches the (uint64_t)PyLong_AsLongLong(value) branch at src/encode.c:1049-1052. For ActiveArea that is a 16-byte read from a full 64-bit attacker-chosen address into the output: a source entry 50829 / LONG8 / 1 / <address of a Python bytes object> produced an output ActiveArea of (b'SECR', b'ET-T', b'OKEN', b'-012') in testing.

The comment above blocklist += at TiffImagePlugin.py:1938-1944 already notes that array tags with the wrong length "will result in segfaults"; the blocklist covers only a handful of tags Pillow itself knows, not the ones libtiff knows and Pillow does not. Forwarding of these tags was introduced in a3d45e9cefd0 ("Added custom int and float TIFF tags", 2018-10-25), first released in 5.4.0; releases 2.0.0–3.0.0 forwarded them as scalars too, until 9fa6634f4 ("Partial fix for #1597") limited _save to LIBTIFF_CORE.

Other branches with the same defect

The scalar-integer branch above is one of several places in the same loop where the call shape is derived from the Python value rather than from libtiff's definition. All of the following were confirmed on this commit (stock build crashes and/or ASan reports, stack always memcpy ← libtiff ← ImagingLibTiffSetField src/libImaging/TiffDecode.c:934 ← PyImaging_LibTiffEncoderNew), and all are reachable from a sub-1 KB crafted file followed by an ordinary save():

• Typed tuple → (len, av) for a fixed-count built-in field (src/encode.c:904-994, the SHORT/LONG/SBYTE/SSHORT/SLONG/FLOAT/DOUBLE tuple branches all call ImagingLibTiffSetField(..., len, av)). For libtiff built-in fields with passcount=0 and a fixed writecount — CFARepeatPatternDim 33421 (SHORT×2), BlackLevelRepeatDim 50713 (SHORT×2), ImageLayer 34732 (LONG×2), ActiveArea 50829 (LONG×4), MatrixWorldToScreen/Camera 33305/33306 (FLOAT×16) — libtiff takes the first vararg as the source pointer, so the tuple length becomes the memcpy source address. A perfectly well-formed 206-byte 8×8 PackBits TIFF carrying 33421 / SHORT / 2 / (2, 2) re-saved with a bare im.save(out, "TIFF") dies with SIGSEGV in memcpy (gdb: rsi = 0x2, the tuple length) on every run; so does Image.new("L", (8, 8)).save(buf, "TIFF", compression="packbits", tiffinfo={33421: (2, 2)}), and a JPEG whose EXIF carries the tag saved with exif=im.getexif() into a compressed TIFF. bytes values for built-in fixed-count BYTE tags (e.g. DNGVersion 50706, BYTE×4) crash the same way through the (PyBytes_Size, PyBytes_AsString) call at src/encode.c:860-865. The mirror image for passcount=1 fields: Photoshop 34377 declared SHORT, count 1, value 60000 takes the scalar branch at :1009-1012, libtiff reads 60000 as the count and an unpassed vararg as the pointer, and the re-save succeeds with a 60162-byte output whose tag 34377 holds 60000 bytes of process memory (here C prototype strings from _imaging.so's .rodata).
• bytes value → single char * (src/encode.c:805-815 autodetects bytes as TIFF_ASCII; :1033-1042 passes PyBytes_AsString(value) alone). ImageFileDirectory_v2._setitem stores a BYTE-typed entry as bare bytes and an ASCII-typed one as str (encoded to bytes by _save), so the file's declared type selects this branch.
  • ColorMap (320) declared BYTE with 4 bytes of data in a non-palette PackBits TIFF: 320 is in core_tags[], no type is consulted, the value skips the 768-tuple TIFFTAG_COLORMAP branch () and libtiff reads three and copies shorts from each — a 512-byte over-read of the 5-byte Python bytes buffer, then dereference of two stale argument registers → SIGSEGV on every run of . is not in 's .

Impact

Any service that accepts a TIFF and writes it back out as TIFF through libtiff — which happens with a plain im.save(fp, "TIFF") whenever the uploaded file is compressed, or whenever a compression= other than raw is requested — returns to the uploader, inside one metadata tag of the produced file, an attacker-chosen number of bytes read from a stale pointer in the worker process, deterministically and without retries. Which memory that is depends on the compiler's register allocation for PyImaging_LibTiffEncoderNew: in this clang build the pointer is _imaging.so + 0x1588, so a request for 262 144 bytes (which succeeded every time in testing, as did 2 000 000) covers the extension's .rodata, .data.rel.ro/GOT and .data and returned 2 800+ resolved pointers into libpython, libc, libtiff, libjpeg and others — a complete ASLR bypass for chaining with any other memory bug; a gcc build of the same source left a different, unrelated address in that slot. The fixed-count variant kills the worker process with SIGSEGV from a 181-byte upload on every request. With tiffinfo=im.tag_v2 the source address is fully attacker-chosen (16-byte arbitrary read into the output for ActiveArea, or a crash). No non-default Pillow setting is involved; the only precondition is TIFF-to-TIFF re-encoding.

The other branches add further deterministic process-kill inputs (BYTE-typed ColorMap in any RGB or palette TIFF; thumbnail() keeps tag_v2, so open→thumbnail→save-as-TIFF is affected too) and extend the reach to JPEG/PNG/WebP uploads whose EXIF is preserved with exif= when converting to libtiff-compressed TIFF. All the memory errors are reads; libtiff sizes its destination buffers with the same count it copies, so no out-of-bounds write was found.

Reproducing

repro is one Python script, run against the stock build with the installed PIL. It needs no compiler and no extra files.

upload(tag, value) has Pillow write a 16×16 red RGB image as an LZW-compressed TIFF, then appends a new IFD holding the same entries plus one extra <tag> / LONG / count 1 / value (kept in tag order) and points the header at it — a 356-byte input on this tree. The source is LZW-compressed on purpose. A bare im.save(out, "TIFF") keeps the source's compression, so the save goes through the libtiff writer even though the application passes no arguments. resave() is the whole "application": Image.open(BytesIO(src)); im.save(out, "TIFF").

The script runs three cases:

1. Control, tag 65000 = 4096. libtiff does not know this tag, so the file is re-saved correctly and the output tag reads back as the integer 4096. This is what should happen to every tag.
2. Leak, DNGPrivateData 50740 = 65536. libtiff's built-in definition of this tag is a counted BYTE array. It reads Pillow's single scalar as the count and takes the data pointer from a vararg slot Pillow never wrote. The output tag comes back as bytes of length 65536 that do not occur in the input. The script prints the first 32 of them and the file offset at which they occur in _imaging.so on disk. On this build they always do (offset 0x57588): the stale pointer aims into the extension's read-only data.
3. Wild read, ActiveArea 50829 = 64. libtiff defines this tag as a fixed LONG[4] array with no count, so it uses the scalar 64 as the source address. The script re-runs itself as a child process (repro child, core dumps disabled) to do this re-save, and the child exits with -11, meaning SIGSEGV.

Expected output on the affected tree:

control  tag 65000=4096 -> 4096
leak     tag 50740=65536 from a 356-byte upload -> 65536 bytes not present in input, head a661fcffc661fcff3360fcff7b61fcffd361fcff085ffcffe061fcff3360fcff
         leaked bytes found in _imaging.so at file offset 0x57588
crash    tag 50829=64 -> child exit -11 (SIGS