CVE-2026-106117

When decoding a fax-compressed strip TIFF (Compression=3 / Group 3 1D, or Compression=2 / Modified Huffman), the CCITT decompressors write decoded runs through BitWriterUtils.WriteBits/WriteBit/WriteZeroBit, which advance and write bits via Unsafe.Add with read-modify-write semantics, without ever comparing the write position against the target buffer length. The strip buffer is sized ImageWidth × RowsPerStrip (8 bytes in the PoC), but two independent defects let an attacker write tens of millions of bits past it: (a) T4 only increments rowsWritten when an EOL code is read, and one ReadNextRun accumulates unlimited makeup codes (+2560 px per 12-bit code) into a single RunLength that WritePixelRun then writes in one shot; (b) Modified Huffman writes before validating (pixelsWritten > Width is checked at :90-93, after the write at :56-63), so the overflow completes even though an exception is thrown later. One crafted ~90 KB file writes ~19.2 MB linearly past an 8-byte buffer and deterministically kills the process; the write offset and length are fully attacker-controlled (classic heap-corruption primitive on the managed heap).

Verified at commit 5cd4d0d26a82a9549f297a237aea9cf665bddff8 (main; latest release v4.1.0, the supported major). A related tiled-path variant (tile-buffer width mismatch) was reported separately as GHSA-v76p-62qx-wwq2, this report covers the distinct strip-path root cause.

Details

Root cause: the bit-writing sink has no bounds check, and neither decompressor constrains run lengths against the strip buffer.

• Unchecked write primitive: BitWriterUtils.cs#L51 (WriteBit), :58 (WriteZeroBit, also read-modify-write, so even all-white runs really write), :11 (WriteBits)
• T4 trigger: T4TiffCompression.cs#L75 and L109-L119, rowsWritten only increments on EOL; consecutive makeup codes accumulate into one RunLength, then WritePixelRun writes it in full
• MH trigger: ModifiedHuffmanTiffCompression.cs#L56-L63, write happens before the width check at L90-L93
• Buffer size: TiffDecoderCore.cs#L932-L967, CalculateStripBufferSize = width × bpp/8 × rowsPerStrip

Attack surface: Image.Load(stream) on an attacker-supplied strip TIFF (TiffDecoderCore.DecodeStripsChunky → TiffDecompressorsFactory.Create → Decompress). Default configuration, no authentication, no user interaction, a plain strip TIFF (far more common than the tiled variant) with Compression=2 or 3 and a chain of CCITT makeup codes.

Relationship to GHSA-v76p-62qx-wwq2: that report is the tiled variant, a caller-side width mismatch (TiffDecompressorsFactory drops tile parameters) that overflows with perfectly legal run codes. This report is the strip variant, the callee-side missing bounds check plus T4's EOL-only row accounting and MH's write-before-validate. Fixing the caller mismatch does not address this vector; bounding BitWriterUtils addresses both (see remediation).

Suggested remediation:

1. Bound BitWriterUtils.WriteBits/WriteBit/WriteZeroBit against buffer.Length*8 (return bool / throw ImageFormatException), do not rely on caller discipline.
2. In the T4/MH decompress loops, validate (bitsWritten + RunLength) <= buffer.Length*8 before writing; abort T4 when accumulated rows exceed stripHeight instead of waiting for the loop to end naturally.
3. In Modified Huffman, move the pixelsWritten > Width check before the actual write.
4. Regression fuzz cases: Compression=2/3, EOL-less oversized makeup chains, edge widths; strip and tiled paths share the same constraint.

PoC

Full PoC posted as the first comment below: Program.cs (driver), poc-tiff-t4.tif + poc-tiff-mh.tif (crafted files, ~90 KB each, base64 inline), README.

1. Build a console project referencing src/ImageSharp/ImageSharp.csproj, run it against either crafted file (or call Image.Load on it).

2. PoC layout: TIFF (II, 42), ImageWidth=64, ImageLength=1, BitsPerSample=1, Photometric=WhiteIsZero, RowsPerStrip=1; strip data = EOL(12bit) + 60000× white makeup 2560 (000000011111) + white terminating code (000111), declared run ≈ 153,600,001 px ≈ 19.2 MB into an 8-byte buffer.

3. Observed (both variants):

   strip payload 90003 bytes, claimed run ~153,600,001 px = 19,200,000 bytes into 8-byte buffer
   Fatal error. System.AccessViolationException: Attempted to read or write protected memory.
      at SixLabors.ImageSharp.Formats.Tiff.Compression.BitWriterUtils.WriteBits(Span`1<Byte>, IntPtr, IntPtr, Byte)
      at ...ModifiedHuffmanTiffCompression.Decompress(...)
      at SixLabors.ImageSharp.Formats.Tiff.TiffDecoderCore.DecodeStripsChunky[Rgba32](...)
   Aborted (core dumped); exit=134
   

   The T4 variant crashes identically at T4TiffCompression.WritePixelRun → BitWriterUtils.WriteBits.

4. Control: an equivalent file without the makeup chain (legal EOL-delimited rows) decodes normally, the crash comes from the oversized runs, not the container.