Model Intelligence Sheet
celvexgroup/modelaudit-gguf-array-chat-template-poc overview
PoC modelaudit GGUF array typed chat template SSTI detector bypass BL 0111 Coordinated disclosure PoC for huntr Protect AI MFV . Benign only. Scanner: modelaud…
Runs locally from ~0.0 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
Model Details
| Model ID | celvexgroup/modelaudit-gguf-array-chat-template-poc |
|---|---|
| Author | celvexgroup |
| Pipeline | — |
| License | — |
| Base model | — |
| Last modified | 2026-07-19T00:56:37.000Z |
Model README
PoC - modelaudit GGUF array-typed chat_template SSTI detector bypass (BL-0111)
Coordinated disclosure PoC for huntr (Protect AI MFV). Benign only.
- Scanner: modelaudit==0.2.49 (gguf_scanner.py). Format: GGUF (.gguf).
modelaudit scan variant_array.gguf-> Clean / exit 0: the tokenizer.chat_template is stored as an ARRAY-of-string (GGUF value_type 9) carrying a plaintext os.popen SSTI, which the Jinja2 SSTI detector misses.control_string.gguf(byte-identical payload as a STRING) fires CRITICAL - proving the detector is live. benign_string.gguf is the negative baseline.- Root cause: gguf_scanner.py:538 gates the chat_template check on
isinstance(value, str), so an ARRAY value (a list) is never recorded into the templates dict fed to the Jinja2 scanner. Distinct from the earlier GGUF finding (which obfuscated inside a string template); this is a metadata value-TYPE confusion, no obfuscation. - Benign: the SSTI payload runs
idand writes a marker. Loader link anchored to CVE-2024-34359 (llama-cpp-python jinja render).
Files: variant_array.gguf (bypass), control_string.gguf (fires), benign_string.gguf. Full report in huntr submission.
Run celvexgroup/modelaudit-gguf-array-chat-template-poc with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models