CVE-2026-70638 - llama.cpp LLaMA-Android JNI Integer Overflow

Companion PoC for the Hunt-Benito article "One Multiply Too Many: CVE-2026-70638 — Integer Overflow in llama.cpp's Android JNI Heap Allocation" → https://www.hunt-benito.com/blog/one-multiply-too-many-cve-2026-70638-llama-cpp-android-jni-integer-overflow/

The bug

Java_android_llama_cpp_LLamaAndroid_new_1batch() in examples/llama.android/llama/src/main/cpp/llama-android.cpp (llama.cpp builds b1886–b7445) is a hand-copy of llama_batch_init() that performs several unchecked malloc(sizeof(T) * count) computations. With an attacker-controlled multiplier (n_seq_max, n_tokens, or embd), the size wraps, the heap block is undersized, and the caller's subsequent writes overflow it (CWE-190 → CWE-122). NVD CVSS 3.1 7.8 High (AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H).

The vulnerable core of the function:

batch->seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);

Files

Run

# 1. Arithmetic demonstrator (the core of the article's PoC)
cc -O2 -o overflow_demo overflow_demo.c
./overflow_demo

# 2. Malicious model file
python3 craft_gguf.py 0x40000001 malicious.gguf   # -> 416-byte GGUF
xxd malicious.gguf | head                           # verify magic 'GGUF' + key

# 3. Live hook (rooted/emulator device with frida-server, authorized target only)
frida -U -l hook_new_batch.js -f <package> --no-pause

Expected demo output (excerpt)

[2] Malicious: n_seq_max = 0x40000000 (2^30). 4 * 2^30 = 2^32 -> wraps:
  n_seq_max (attacker)   = 1073741824 (0x40000000)
  malloc size, arm64     = 4294967296 bytes (4.00 GiB)
  malloc size, armeabi-v7a = 0 bytes
  caller believes it got = 4294967296 bytes
  >>> 32-bit WRAP: 4294967296-byte write into 0-byte heap block (CWE-122)

Scope & accuracy notes

• The wrap-to-undersized-allocation is directly reachable on 32-bit ABIs (armeabi-v7a, still shipped). On 64-bit (arm64-v8a) the same unchecked expression instead requests a multi-GiB block that malloc fails (NULL → later crash / DoS); heap corruption requires the 32-bit path.
• overflow_demo.c performs no out-of-bounds write — it only proves the arithmetic wraps. craft_gguf.py produces a non-runnable model (no tensors) so it cannot be weaponized as-is; it demonstrates that the attacker-controlled multiplier is sourced from untrusted metadata.

Remediation

Upgrade llama.cpp to b7446 or later (the new_1batch JNI path was removed by the Android-binding rewrite). If you must stay on an affected build or a downstream fork, apply Cyera's validation patch from https://github.com/Vladimir-tokarev-cyera/llama-cpp-security-patches (the CVE-2026-43627 batch-init guard is the canonical fix for this bug class).

Sources

• NVD — https://nvd.nist.gov/vuln/detail/CVE-2026-70638
• Fix (rewrite) — https://github.com/ggml-org/llama.cpp/commit/5c0d18881e0e9794c96b2602736b758bac9d9388
• Affected source — https://github.com/ggml-org/llama.cpp/blob/b7445/examples/llama.android/llama/src/main/cpp/llama-android.cpp

FOR AUTHORIZED SECURITY RESEARCH AND EDUCATIONAL USE ONLY.