darknet-poc

⚠️ SECURITY RESEARCH ARTIFACT — DO NOT LOAD

This repository contains a deliberately malicious Darknet model pair (.cfg + .weights) used as the proof-of-concept for a coordinated-disclosure vulnerability report filed via huntr.com against hank-ai/darknet (the maintained Darknet / YOLO native framework).

Do NOT load these files with darknet or any Darknet-based tool on a machine you care about. Loading b1_e2e.weights triggers a heap out-of-bounds write during weight transposition and will corrupt process memory / crash the loader.

What it triggers

• Weakness: CWE-190 (Integer Overflow) → CWE-787 (Out-of-bounds Write), heap.
• Sink: transpose_matrix() in src-lib/weights.cpp, reached at model-load time via load_weights → load_weights_upto → load_connected_weights → transpose_matrix.
• Mechanism: the .weights header sets major=1001, which enables the transpose path. The [connected] layer's inputs (65536) × outputs (65537) overflows the 32-bit int used to size the transpose scratch buffer (wraps to 65536), so xcalloc under-allocates, but the transpose loop iterates the true rows × cols range and writes past the buffer.
• Verified end-to-end under AddressSanitizer against the real framework parser + loader: heap-buffer-overflow WRITE of size 4, 0 bytes after a 262144-byte region in transpose_matrix.

Files

Note on file size: the .weights file carries the full ~512 KiB float payload on purpose. Darknet's xfread fatally aborts on a short read, so the biases and weights reads must fully succeed for execution to reach the transpose sink. A truncated file dies before the bug. The b1_e2e.cfg prepends a no-op 1×1 [convolutional] layer solely to preserve the 256×256×1 input volume (feeding inputs=65536) and to avoid an unrelated divide-by-zero on a connected-only network — see the huntr report for details.

Disclosure status

Filed for coordinated disclosure through huntr. Please do not redistribute or public-post this PoC until the maintainer coordination and disclosure window have completed. This artifact exists only to let the maintainers and huntr triagers reproduce and fix the issue.

Contact: huntr handle Ano1X8.