Security context

What an agent needs to avoid regressing past fixes and find the next vuln in this repo.

ai-hypercomputer/maxtext

              main @ 442c205
            

Fixes

CVEs

HIGH

Peak severity

25.0%

Coverage

Highlights

Code Execution: 1 prior fix. Scrutinize any change in this area.

scripts/utilities: most-fixed (1 issue). Treat as high-risk during review.

1 high-severity fix in this history; regressions here are high-impact.

Recurring patterns

The bug types that recur here, drawn from past fixes, not open vulnerabilities.

Code Execution: The checkpoint loading pipeline used insecure deserialization via torch.load, allowing arbitrary code execution when loading untrusted model weights. Developers must enforce weights_only=True and avoid unsafe pickle formats.

Information Disclosure: Obsolete and vulnerable JAX Pallas Paged Attention kernels and page manager allocators leaked cross-tenant HBM memory due to improper isolation of memory allocation.

Denial of Service: Negative prefetch indexing in TPU ragged attention kernels triggered TPU crashes due to unclamped index divisions. Values must be clamped to prevent negative indexing.