AI Security Starts below the model
Episode details
Episode description
61% of enterprises running AI agents share credentials across them — and only 14% can revoke an agent's access within minutes if something goes wrong. This is the infrastructure security debt that's accumulating underneath every AI deployment. Nir Chervoni, Head of Security Products at Nebius, lays out exactly what must be locked down before a model or agent ever runs: per-agent identity scoped like a human user account, short-lived credentials stored in a secrets manager (not prompts or env vars), encryption keys under customer control, and audit trails feeding directly into SIEM. His foundational rule for architects: never rely on a guardrail to enforce something an access policy could enforce deterministically. The conversation goes deeper into GPU confidential computing — a shift that extends the trust boundary from CPU to GPU, encrypts the CPU-to-GPU data path, and enables attestation so customers can verify the environment before releasing their keys. Chervoni is direct about what it doesn't cover: it won't stop prompt injection, won't fix application vulnerabilities, and is only as strong as whoever verifies the attestation. On sovereign AI, Chervoni draws a critical distinction: data residency answers one question (where data sits), while sovereignty answers three more — who can access it, how it's processed, and who controls the protections. Most enterprises stop at residency and miss the third question entirely. For CISOs evaluating AI-specialized cloud providers, he closes with a checklist of non-negotiable basics: GPU memory scrubbing between tenants, isolated management networks, per-tenant InfiniBand or RoCE fabric partitioning, and third-party pen testing of tenant isolation specifically. GUEST: Nir Chervoni | Head of Security Products, Nebius HOST: Rob Strechay | VentureBeat Intelligence --- **Chapters** 00:00 Intro: The Agent Credential Problem — VentureBeat Pulse Data 01:20 What Must Be Secured Before a Model or Agent Runs 02:00 Infrastructure Controls vs. Guardrails vs. Application Controls 03:20 GPU Confidential Computing: What Changes at the GPU Layer 04:40 Limits of Confidential Computing — What It Doesn't Cover 05:40 Cross-Tenant Risks in GPU Clouds: What…