ort 2.0.0-rc.15 is a memory-safety release, and pyke says to upgrade now

pyke published ort 2.0.0-rc.15 on 19 October 2026, the release after rc.14 made the ONNX Runtime environment explicit. This one has no new features. In the maintainer’s words, “rc.15 fixes over 30 instances of undefined behavior”, and “It is highly recommended to upgrade to rc.15 as soon as possible, as multiple memory leaks/memory safety issues were fixed.”

ort is the widely used Rust binding for Microsoft’s ONNX Runtime, so this reaches anyone running ONNX models from Rust. The crates.io record shows the version published at 07:23 UTC on 19 October, with a minimum Rust version of 1.92.

What breaks when you upgrade

The release notes list six breaking changes:

• run_async now only accepts owned inputs.
• ExecutionProviderLibrary::unregister is removed. You can still unregister execution-provider libraries through the unsafe -sys API.
• ValueRefMut no longer dereferences mutably to Value. pyke recommends taking ValueRefMut wherever you used to take &mut Value.
• InMemorySession no longer dereferences mutably to Session.
• Kernel now requires Send, and ThreadManager requires Send + Sync.
• SessionBuilder::with_external_initializer_file_in_memory takes Arc<[u8]> instead of Cow<'static, [u8]>.

The reason for the first change is spelled out in its commit message: borrowed inputs were a problem because “on drop, onnxruntime could continue to access an input’s now freed memory.” That is the pattern behind most of this release: API shapes that let safe Rust code hand ONNX Runtime memory that could go away underneath it.

The compare view shows 37 commits across 65 files between the two tags.

Advisory or not

The repository has no published security advisories for these fixes. So describe this as memory-safety fixes, not as numbered vulnerabilities. We also have not mapped each fix to the version that introduced the bug, so do not assume every older release is affected by every one.

The notes also say “Alternative backends will be updated sometime before Oct 23.” If you use one of ort’s alternative backends rather than ONNX Runtime itself, check for that update before you upgrade.

A note for Mac users

Issue #662, opened on 13 October, reports that on rc.14, registering the CPU execution provider does not override the default policy, so sessions run on CoreML on macOS. It was still open when we checked on 19 October. If your models behave differently or run slower on a Mac after moving past rc.13, look there first.

What to do

1. Bump ort to 2.0.0-rc.15 and run cargo build. The compiler will point at most of the breaking changes.
2. Replace borrowed run_async inputs with owned values, and swap &mut Value parameters for ValueRefMut where needed.
3. Rerun your inference tests. These fixes target memory handling. A release-candidate bump is still worth a full test pass.

ort is still a release candidate. The 2.0 final release has not shipped.