numpy-npy-poc

numpy/numpy — Uncaught tokenize.TokenError in np.load()

Summary

Loading a crafted 1,214-byte NPY file causes np.load() to raise tokenize.TokenError instead of the documented ValueError. The exception escapes uncaught from all numpy handlers, crashing any application that catches only ValueError per numpy's documented API.

Target Info

Root Cause (3 lines)

_read_array_header() calls _filter_header() when ast.literal_eval() raises SyntaxError. For headers with 100+ nested parentheses, tokenize.generate_tokens() inside _filter_header() raises tokenize.TokenError — which is not a subclass of ValueError and is never caught.

# numpy/lib/_format_impl.py line 661
except SyntaxError as e:
    if version <= (2, 0):
        header = _filter_header(header)  # <-- raises TokenError, not caught!

Trigger Path

crafted NPY v1.0 (100-level nested descr)
  → np.load()
  → format.read_array()
  → _read_array_header()
  → ast.literal_eval()  raises SyntaxError
  → _filter_header()
  → tokenize.generate_tokens()  raises TokenError
  → [UNCAUGHT] propagates out of np.load()

Reproduction

# 1. Install numpy (tested on 2.3.5 + Python 3.13)
pip install numpy

# 2. Run PoC
python3 poc_numpy.py

# Expected output:
# [+] CONFIRMED: tokenize.TokenError raised!
# [+] CONFIRMED: TokenError ESCAPED ValueError handler!
# depth=99: ValueError raised (safe)
# depth=100: [+] TokenError raised (VULNERABLE)

PoC File Stats

Suggested Fix

# In _read_array_header(), wrap _filter_header() call:
from tokenize import TokenError as _TokenError

try:
    header = _filter_header(header)
except _TokenError as te:
    raise ValueError(
        f"Cannot parse header (tokenizer error): {header!r}"
    ) from te

Files