rules-numpy

RISK: np.load(allow_pickle=True) deserializes arbitrary Python objects from the file, enabling remote code execution if the .npy file comes from an untrusted source (CWE-502, analogous to pickle exploits). FIX: use allow_pickle=False (the default since NumPy 1.16.3) and store structured data as .npz (compressed arrays) or .csv/.json for human-readable formats.

RISK: np.matrix is deprecated since NumPy 1.15 and slated for removal — it wraps ndarray with different broadcast semantics that silently change the result of * (matrix multiply vs element-wise). FIX: replace with np.ndarray (np.array) and the @ operator for matrix multiplication.

RISK: np.frompyfunc(eval, ...) vectorizes the Python eval function over array elements — passing an array of untrusted strings causes code execution on every element (CWE-94). FIX: never pass eval/exec/compile to frompyfunc; use np.vectorize with a safe parsing function.

RISK: constructing string arrays from np.fromstring/frombuffer and later evaluating them (e.g., with eval or ast.literal_eval in a loop) creates a code-execution surface if the data is user-controlled. FIX: validate input data against an expected schema and use structured array dtypes instead of raw string buffers.

RISK: dtype=object arrays store arbitrary Python objects — serializing them with pickle (np.save/joblib) is equivalent to raw pickle and enables deserialization RCE when loaded from untrusted sources (CWE-502). FIX: use a concrete numeric dtype (float64, int32) and ensure object arrays are never persisted or loaded from external data without schema validation.