Descriptions:
Lavina D’Mello, a senior software developer on NVIDIA’s deep learning infrastructure team, makes the case at AI Engineer that the most dangerous security failures in production ML aren’t sophisticated AI-native attacks — they’re the same infrastructure misconfigurations the industry believed it had solved years ago. Her talk’s provocation: the LLM stack is being secured with a 2008 playbook applied to a 2026 system.
D’Mello opens with the 2023 Ray cluster incident, where thousands of distributed ML clusters were exposed on the open internet because authentication was off by default — a lapse resulting in over $1 billion in exposed compute. She cites independent research auditing 50 real production ML setups: 78% had at least one critical security mistake, consistently around open access controls, absence of workload isolation, and credentials stored in publicly accessible object storage.
Her four-layer defense framework — infrastructure, access control, runtime security, and organizational culture — comes with explicit latency cost data. Basic controls cost under 8% overhead and should always be on; workload isolation runs 10–20% and should be applied selectively; real-time malicious input detection costs 15–30% and is often impractical at request scale. D’Mello’s through-line: implementation quality determines whether a security layer is viable, and cultural accountability is the foundation on which every technical control rests.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Deep Dive
