About this session

ShadowRay exposed over a billion dollars of data through a missing authentication check. It wasn't a zero-day. It wasn't a clever new attack class. It was a default config someone never flipped off. That story is not the exception in production ML, it's the rule. We synthesized 139 peer-reviewed papers on production ML security across access control, runtime security, infrastructure, and operations. Five findings stood out, and one of them upends how most teams think about ML security: - Misconfiguration, not missing features, is the dominant failure mode. The mechanisms exist. Teams aren't using them, or are using them wrong. - Adversarial defenses impose 15–30% inference overhead, which is why almost no production system actually runs them. - ML-specific security tooling lags general DevOps tooling by years. - Security, data-science, and ops teams operate in expertise silos that create persistent gaps no single team can see. - LLM and multi-tenant GPU threats are evolving faster than defenses (prompt injection, RAG poisoning, GPU side channels). This talk walks through the four-pillar defense-in-depth framework, the six-category threat taxonomy that maps each attack to its primary and secondary defenses, and a four-level security maturity model that matches overhead budgets to deployment contexts. You leave knowing where your stack actually sits and which 3 misconfigurations account for most of the risk.

Topics

Speaker

Lovina Dmello

Senior Software Developer · NVIDIA