An opaque model is a machine learning model whose internal details, training lineage, or provenance are hard to inspect or verify. Defenders may not know exactly which datasets shaped it, what version is running, whether it was modified, or how faithfully it matches the approved artifact. That lack of visibility makes it harder to assess risk, reproduce behavior, or prove integrity.
Opaque models matter in cyber security because attackers can hide malicious changes, smuggle in unvetted models, or exploit gaps in trust and governance. In defense, teams use inventories such as AI-BOMs, model signatures, and provenance records to reduce opacity and detect tampering or shadow deployments. Still, documentation alone does not prove a model is safe: an opaque model can be signed, labeled, or packaged neatly while remaining difficult to audit. Security improves when transparency is paired with validation, review, and runtime monitoring.