0xiviel/poc-coremltools-rce

PoC: CoreMLTools Arbitrary Code Execution via Malicious Sensitivity File

Vulnerability

Apple's CoreMLTools library calls torch.load() without `weights_only=True` at three locations in sensitive_k_means.py (lines 540, 549, 649).

The SKMPalettizer.compress(sensitivity_path=...) method accepts a user-controlled file path and loads it via torch.load() without the safety flag, allowing arbitrary code execution via pickle deserialization.

Files

• —malicious_sensitivity.pt — Crafted PyTorch checkpoint that executes id when loaded (305 bytes)
• —create_malicious_pt.py — Script to generate the malicious .pt file
• —poc_coremltools_rce.py — Full PoC demonstrating the vulnerability

Reproduction

# Requires: pip install torch coremltools
python3 poc_coremltools_rce.py

# Or test just the torch.load part (no coremltools needed):
python3 poc_coremltools_rce.py --torch-only

Impact

• —Arbitrary code execution when loading a crafted .pt file as sensitivity values
• —Attack scenarios: shared ML training environments, CI/CD pipelines, tutorials with malicious downloads

NON-DESTRUCTIVE

This PoC only executes id to demonstrate code execution. No files are modified or deleted.