Security: Fix CWE-95 (Eval Injection) in _get_torch_dtype()

[yoavkatz]

Security Fix: CWE-95 (Eval Injection)

Summary

Fixed a HIGH severity security vulnerability (CVSS 7.8) in the _get_torch_dtype() method that could allow arbitrary code execution.

Vulnerability Details

• CWE: CWE-95 (Eval Injection)
• Severity: HIGH (CVSS 7.8)
• Location: src/unitxt/inference.py:550 (before fix)
• Reported by: External security researcher via IBM PSIRT

The _get_torch_dtype() method used Python's eval() function with insufficient input validation, which could be exploited to execute arbitrary code.

The Fix

Replaced the eval() call with a secure lookup table containing an explicit whitelist of 21 valid torch dtypes:

torch_dtypes = {
    "torch.float16": torch.float16,
    "torch.float32": torch.float32,
    "torch.bfloat16": torch.bfloat16,
    # ... 18 more valid dtypes
}
dtype = torch_dtypes.get(self.torch_dtype)

Changes

• src/unitxt/inference.py: Replace eval() with explicit whitelist of valid torch dtypes
• tests/inference/test_inference_engine.py: Add comprehensive security tests
  • test_torch_dtype_security_fix() - Full integration test
  • test_torch_dtype_security_fix_fast() - Fast unit test (1.7s)

Security Improvements

✅ Eliminates code injection vulnerability
✅ Validates input against explicit whitelist
✅ No breaking changes - all legitimate torch dtypes continue to work
✅ Better performance (dict lookup vs eval)
✅ Clearer error messages listing supported values

Testing

All tests pass including new security-specific tests:

pytest tests/inference/test_inference_engine.py -k test_torch_dtype_security -v
# PASSED

Impact Assessment

• ✅ No breaking changes
• ✅ No user action required
• ✅ Performance improvement
• ✅ Better UX with clearer error messages

Security Note

Detailed exploitation information has been shared privately with the security team and will be disclosed responsibly after the fix is released.