Model Intelligence Sheet
hacnho/gguf-add-bos-token-poc overview
GGUF add bos token PoC This repo is a benign security research PoC for Huntr MFV triage. Files: command r no bos.gguf verify remote poc.py Tested trigger entry…
Runs locally from ~10.4 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
1 GGUF files detected
Direct downloads for local inference
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| command-r-no-bos.gguf | GGUF | GGUF | 10.4 MB | Download |
Model Details
Model README
GGUF add_bos_token PoC
This repo is a benign security research PoC for Huntr MFV triage.
Files:
command-r-no-bos.ggufverify_remote_poc.py
Tested trigger entrypoint:
llama-tokenize -m <model> --ids -p <prompt>
Behavior:
- seed vocab-only GGUF uses
tokenizer.ggml.add_bos_token = true - malicious artifact flips that metadata to
false - ordinary prompt tokenization then silently drops the BOS token
Reproduction
python3 verify_remote_poc.py --repo hacnho/gguf-add-bos-token-poc
Expected output delta
Prompt:
hello world
Control token IDs:
[5, 34313, 3845]
Malicious token IDs:
[34313, 3845]
The same BOS-dropping behavior also appears on:
hello worldhello
Scanner posture
Local scout observed:
modelscan:No issues foundon the malicious artifact
Run hacnho/gguf-add-bos-token-poc with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models