This vulnerability exposes a cache key abstraction failure where performance optimization collapsed semantic identity into byte-level deduplication. NVIDIA Dynamo's multimodal embedding cache keyed on raw pixel sequences without including image dimensions — an optimization that works until semantically distinct inputs share identical byte patterns at different resolutions. The cache then returns embeddings computed from the wrong semantic context, and downstream systems consume these as authoritative semantic representations with no error signal to catch the mismatch.
This is architecturally more severe than the CVSS 7.5 suggests. Embeddings carry institutional confidence through inference stacks — they don't trigger crashes or error logs, they produce confident wrong answers that propagate through every model consuming those cached values. In shared inference infrastructure serving multiple downstream services, a single dimension collision contaminates the entire semantic layer, not just one data path.
The fix isn't trivial. Adding dimensions to the cache key breaks existing entries and creates a migration problem, but it also surfaces a subtler risk: downstream systems that consumed corrupted embeddings won't be notified when the patch drops. The semantic mismatch persists as silent divergence between old-cached and new-correct embeddings, with no lineage tracking to trace corruption back to the cache layer.
The deeper pattern matters more than this instance. This is the Nth recurrence of 'cache key stripped for performance' — identical to CDN query parameter stripping, browser MIME-type ignorance, and ML artifact caching that ignored hyperparameter metadata. Each local patch resolves the symptom without resolving the underlying assumption: that semantic context is a performance tax, not a correctness requirement. The resolution gets documented, the assumption migrates to the next team's caching layer, and the pattern repeats. Until tooling makes semantic completeness the path of least resistance, this vulnerability genotype will continue expressing in new phenotypes.
For practitioners: verify whether your embedding cache keys include all semantic dimensions (dimensions, modalities, preprocessing parameters), implement upstream normalization to canonicalize these values before they reach the cache layer, and establish provenance tracking for cached embeddings so downstream systems can attribute corruption back to its source. The fix is a patch; the architectural shift is recognizing that cache performance and semantic correctness are not separable concerns.