You can deploy this model and start earning money today!

jdhart81/
llamafile-poc-llf002

Name: llamafile-poc-llf002
Rating: 2.5 (9 reviews)

llamafile

Views

No views yet

LLF-002: llamafile ZIP Central Directory Out-of-Bounds Read

Target: Mozilla-Ocho/llamafile File: llamafile/llamafile.c, lines 155-175 (ZIP central directory parsing) Severity: MEDIUM (CWE-125 Out-of-Bounds Read) Reporter: Viridis North LLC | viridisnorthllc@gmail.com

Summary

A crafted .llamafile ZIP container can cause out-of-bounds heap reads during central directory parsing. The EOCD (End of Central Directory) record count field is trusted without validation against the actual central directory size, causing the parsing loop to iterate past valid entries into uninitialized heap memory.

Vulnerable Code

// llamafile.c:155-175
cnt = ZIP_CDIR_RECORDS(eocd);     // from EOCD — UNTRUSTED
cdirsize = ZIP_CDIR_SIZE(eocd);   // from EOCD
cdirdata = malloc(cdirsize);
pread(fd, cdirdata, cdirsize, off);

for (int i = 0; i < cnt; i++) {   // iterates 'cnt' times
    // entry_offset + HDRSIZE check passes for first real entry
    // but after exhausting real entries, reads garbage from heap
    if (ZIP_CFILE_MAGIC(cdirdata + entry_offset) != kZipCfileHdrMagic)
        goto next;  // checks magic — but reads from uninitialized memory first
    // ... memcmp, strndup on potentially OOB data
}

The cnt field from EOCD declares 100 records, but only 1 actual central directory entry exists (56 bytes). After the first valid entry, the loop reads ZIP_CFILE_MAGIC from whatever follows in the malloc'd buffer — uninitialized heap memory.

Attack Vector

User opens a crafted .llamafile file. The ZIP container has a manipulated EOCD record count. During model loading, the central directory parser reads out of bounds, potentially leaking heap contents or crashing.

PoC

llf002_excess_cdir_records.llamafile (142 bytes):

Contains 1 valid GGUF model entry in the central directory (56 bytes)
EOCD declares cnt=100 records
After parsing the 1 real entry, 99 more iterations read from uninitialized heap

Generate PoC (Python)

python

import struct, zlib
gguf = struct.pack('<II', 0x46554747, 3) + struct.pack('<QQ', 0, 0)
name = b'model.gguf'
# Local file header + Central directory + EOCD with cnt=100
# See full generator in repo

Test

bash

# Build with ASan
make -j$(nproc) CXXFLAGS="-fsanitize=address,undefined -O1 -g" LDFLAGS="-fsanitize=address,undefined"

# Run
ASAN_OPTIONS=detect_leaks=0 ./llamafile --model llf002_excess_cdir_records.llamafile 2>&1
# Expected: heap-buffer-overflow or use-of-uninitialized-value

Impact

CWE-125: Out-of-bounds read from heap
Heap content disclosure (information leak) or crash (DoS)
The uninitialized memory may contain sensitive data from prior allocations

Suggested Fix

Validate cnt against cdirsize before the loop:

if (cnt > cdirsize / kZipCfileHdrMinSize) {
    cnt = cdirsize / kZipCfileHdrMinSize;  // cap to maximum possible entries
}

jdhart81/llamafile-poc-llf002