Mamba EUVD-2026-29562

Primary remediation requires upgrading the Mamba framework to a patched version that implements weights_only=True in torch.load() calls within from_pretrained() methods-however, no vendor-released patch version is confirmed from available data at time of analysis; monitor the official GitHub repository (github.com/state-spaces/mamba) and PyPI for security updates. Immediate compensating control: implement model provenance verification by only loading models from explicitly whitelisted, organization-controlled repositories; cryptographically verify model file hashes against known-good values before loading; this prevents execution but adds operational overhead for model updates. Advanced mitigation: monkey-patch torch.load in application initialization to force weights_only=True globally (may break legitimate non-tensor pickle data); this could cause compatibility issues with models using custom serialization. Containerization mitigation: run Mamba model loading in isolated containers with minimal privileges and no network access post-download; limits blast radius but does not prevent initial compromise. Long-term: migrate to safer serialization formats like safetensors which stores only tensor data without pickle vulnerabilities; requires model re-export and framework support. For security-critical environments, consider disabling HuggingFace Hub integration entirely and distributing models through internal registries with mandatory security scanning.