Microsoft releases ONNX Runtime v1.30.1 with memory safety and model validation fixes

Brief written by the FOSSWire AI editor from ONNX Runtime How we write

October 29, 2026 1 min read

Key Takeaway

This patch release focuses on hardening shape inference and fixing critical memory safety vulnerabilities like heap buffer overflows and out-of-bounds reads.

Summary

Microsoft has released ONNX Runtime version 1.30.1, a patch update following the v1.30.0 release. This version primarily addresses security and stability concerns through targeted model-validation improvements and memory-safety fixes. It is designed to prevent crashes and unauthorized memory access when processing complex AI model graphs.

The update affects developers and system integrators using ONNX Runtime for local or edge AI inference. Key improvements include hardened shape inference for specific operators, fixes for out-of-bounds reads during session initialization, and enhanced data transfer handling for Execution Providers (EP). The release also includes build and packaging refinements for Windows and Python environments.

Technical improvements in ONNX Runtime v1.30.1

Memory Safety: Fixed a heap buffer overflow occurring when packed sub-byte tensor storage was incorrectly reused for larger full-byte tensors, and prevented out-of-bounds reads in EmbedLayerNormalization and ConvTransposeWithDynamicPads shape inference.
Model Validation: Hardened shape inference for contrib-ops including CausalConvWithState, GatedDeltaNet, GroupQueryAttention, and SparseAttention by adding rank, attribute, and overflow validation.
Graph Integrity: Implemented checks to reject malformed graphs containing inputs to zero-input operator schemas or subgraphs located on non-control-flow nodes.
Execution Provider Optimization: Improved Plugin EP single-tensor data transfers by preserving the caller's execution stream, enabling asynchronous copies to utilize the supplied stream while maintaining synchronous-copy behavior.
Build and Packaging: Updated Windows build-tools to utilize system-trusted certificates and refined the Python packaging stage to explicitly select Python 3.12 before validation.

Why it matters

By addressing heap overflows and out-of-bounds reads, this release increases the security posture of AI deployment pipelines, particularly when running untrusted models. The improvements to Execution Provider data transfers and shape inference stability provide a more reliable foundation for developers building cross-platform AI applications on diverse hardware backends.

Original reference: github.com →