CastForge is designed to safely process untrusted model artifacts (.mlir, .safetensors, .gguf, .onnx.json, and .json):

1. Checked Dimension & Byte-Offset Arithmetic: Tensor element counts and byte offsets are validated using checked integer arithmetic (checked_mul / checked_add) with a hard cap of TensorType::MAX_TENSOR_ELEMENTS (1,000,000,000 elements) and a 100 MiB header limit on .safetensors metadata.
2. Bounded Region Recursion Depth: The MLIR parser enforces MAX_REGION_NESTING_DEPTH (64) to prevent call-stack exhaustion on deeply nested regions.
3. Panic-Safe C FFI Boundary: All extern "C" entry points in src/ffi.rs wrap execution in std::panic::catch_unwind to guarantee that panics never cross FFI boundaries.

If you discover a security vulnerability (such as an out-of-bounds read, integer overflow, or unbounded allocation triggered by malformed input), please report it privately via GitHub Security Advisories on the repository rather than opening a public issue.