NVIDIA says production ML keeps failing from misconfigurations, not model exploits
AI Engineer · youtube · 2026-07-21
NVIDIA's Lovina Dmello argues that production ML systems keep getting breached for the same boring reason: they are secured like models instead of like infrastructure.
- She cites a 2023 finding that thousands of Ray clusters were left exposed on the public internet, with dashboards and job APIs reachable because authentication shipped off by default.
- In a review of 139 peer-reviewed papers on production ML security, the dominant failure mode kept recurring as misconfiguration rather than novel exploits.
- One audit of 50 real systems found critical mistakes in 78% of them, usually the same three issues: overprivileged accounts, flat networks, and secrets or model weights left readable in storage.
- Heavy defenses often add 15% to 30% inference overhead, so the practical challenge is deploying controls cheaply enough to survive production.
More from coding & agent
- ty now reads Pydantic config keywords and field metadata — charliermarsh · 2026-07-22
- Pensar Launches AI Security Agent to Autonomously Discover and Patch 0-Days — andriy_mulyar · 2026-07-22
- ty adds first-class Pydantic support, including strict and lax field handling — charliermarsh · 2026-07-22
- Google launches Gemini 3.5 Flash Cyber for CodeMender, with limited access for governments — GoogleAI · 2026-07-22
- Cursor doubles usage limits across all plans for Grok, Composer and new models — XFreeze · 2026-07-22
- Video-based proof of work is emerging as a feedback layer for coding agents — Vjeux · 2026-07-22