README.md

---
tags:
  - security
  - vulnerability
  - poc
  - skops
  - numpy
  - sklearn
  - huggingface
  - decompression-bomb
  - cwe-789
license: mit
---

# skops — .skops NPY Dtype Decompression Bomb (PoC)

**Repo:** `MBM7/skops-dtype-decompression-bomb-poc`  
**Status:** Responsible disclosure — submitted to Huntr  
**Severity:** High / CWE-789  
**Package:** `skops` (PyPI) — HuggingFace official sklearn model format

---

## Summary

A crafted **8.9 KB** `.skops` sklearn model file causes `skops.io.loads()`
to allocate **2 GB of memory** before failing.

| File size | Claimed dtype itemsize | Peak allocation | Amplification |
|-----------|------------------------|-----------------|---------------|
| 8,935 bytes | 500,000,000          | **500 MB**      | 1 : 55,959    |
| 8,935 bytes | 2,000,000,000        | **2,000 MB**    | 1 : 223,838   |

---

## Root Cause

`skops/io/_numpy.py`, `NdArrayNode._construct()`:

```python
content = np.load(
    self.children["content"],   # BytesIO of .npy file from .skops ZIP
    allow_pickle=False
)   # ← np.load allocates count * dtype.itemsize BEFORE reading data
```

The `.skops` format is a ZIP archive containing `schema.json` + `.npy` files.
When loading, `NdArrayNode` reads each `.npy` file into a `BytesIO` and passes
it directly to `np.load()` — **with no validation of dtype.itemsize or shape**.

A crafted `.npy` file inside the ZIP claiming `dtype=|S2000000000` (2 GB
itemsize) causes `np.load()` to pre-allocate 2 GB before attempting to
read data.

---

## Attack

Replace any `.npy` file inside a legitimate `.skops` model archive with a
dtype bomb `.npy` header. `schema.json` remains unchanged — skops trusts it
and loads all `.npy` files unconditionally:

```python
import skops.io as sio

# Load triggers NdArrayNode._construct() → np.load() → 2GB allocation
trusted = sio.get_untrusted_types(data=bomb_skops_bytes)
model = sio.loads(bomb_skops_bytes, trusted=trusted)
```

---

## Distinct from numpy NPY finding

| | numpy NPY | **skops** |
|--|--|--|
| Entry point | `np.load('file.npy')` | `skops.io.loads(skops_bytes)` |
| Format | `.npy` raw | `.skops` ZIP (sklearn model) |
| Ecosystem | NumPy arrays | HuggingFace sklearn model sharing |
| Context | Direct array loading | sklearn model deserialization |

---

## Reproduce

```bash
pip install skops scikit-learn numpy
python poc_skops_dtype_bomb.py
```

Expected:
```
.skops file size  : 8,935 bytes
Expected allocation: 2,000,000,000 bytes
Amplification     : 1:223,838
Result            : ValueError: EOF: reading array data...
Peak memory       : 2000 MB  ← allocated before error
```

---

## Impact

Any service loading user-supplied `.skops` files from HuggingFace Hub or
other sources can be OOM-killed by a sub-10KB payload. skops is the
**official HuggingFace format** for sharing sklearn models — directly
relevant to the AI/ML supply chain.

---

## Suggested Fix

In `NdArrayNode._construct()`, validate before `np.load()`:

```python
MAX_ARRAY_BYTES = 512 * 1024 * 1024  # 512 MB
buf = self.children["content"]
# Peek at NPY header without allocating
hdr = np.lib.format.read_array_header_1_0(buf)
buf.seek(0)
if hdr[1].itemsize > MAX_ARRAY_BYTES:
    raise ValueError(
        f"NPY dtype itemsize {hdr[1].itemsize} exceeds limit — "
        "possible decompression bomb"
    )
content = np.load(buf, allow_pickle=False)
```

---

## Environment

| Package | Version |
|---------|---------|
| skops   | 0.14.0  |
| numpy   | 2.4.4   |
| Python  | 3.12    |

---

*skops was created as a **safer alternative to pickle** for sklearn models.*  
*Ironic: the safety-focused format has an unvalidated allocation path.*