# Securing AI at Scale: CoreWeave and CrowdStrike on What Comes Next

## Executive summary

The conversation addresses the critical challenge of securing AI infrastructure as AI moves into mission-critical roles. Speakers emphasize that AI accelerates the threat landscape, rendering historical security metrics like 'breakout time' obsolete. Solutions require a shift from traditional patching and rules-based security to advanced, multi-layered architectural defenses. Key strategies include leveraging massive compute power to train defensive models (e.g., CrowdStrike SafeMind), achieving deep observability across the entire process and network stack, and implementing rigorous security controls across the entire AI lifecycle, from model development to runtime execution.

## Key takeaways

- Breakout Time is Obsolete: The concept of 'breakout time' (the time from initial adversary access to achieving a goal) has compressed from months/weeks to seconds, making traditional detection windows insufficient. The speed and coordination of modern agents are the primary threat vectors. (Bartley Richardson)
- Security Debt is Existential: The industry faces '50 years of security debt' that is now being attacked by AI-powered adversaries. The focus must be on architectural resilience and automated response to counter this accelerating risk. (Jim Higgins)
- Defense Requires Superhuman Intelligence: Defenders must leverage massive data sets (e.g., 2 billion hours of streaming 4K footage daily) to train defensive models. This process involves red/blue teaming in rich virtualized environments (tens of thousands of VMs) to continuously harden systems. (Bartley Richardson)
- Compute Focus: State, Not Just General Knowledge: AI models must be trained to be 'experts at state'—understanding the operational context of an environment—rather than just general cybersecurity knowledge. This requires constant, scaled training runs (e.g., RL alignment). (Jim Higgins)

## Technical details

- AI Threat Acceleration: Adversaries are increasingly using autonomous agents and swarms, leading to sophisticated, coordinated attacks that exploit misconfigurations and stolen credentials. The threat is characterized by its speed and scale. (Bartley Richardson)
- CrowdStrike SafeMind: A defensive technique built on substantial compute power, SafeMind uses red/blue teaming in highly virtualized environments to continuously test and harden systems against advanced breakout attempts. (Bartley Richardson)
- AI Security Lifecycle: Protection must be multi-layered: 1) Architectural components (DPU, GPU, LPU) for visibility; 2) Model alignment and safeguards; 3) Runtime security (e.g., NVIDIA agent safety platform); and 4) Deep observability into the process and network stack, regardless of whether the activity is human or agent-driven. (Sasha Okonnell)
- Compute Requirements: Training advanced defensive models is highly compute-intensive and requires continuous, scaled training runs (e.g., RL alignment) to maintain efficacy. The goal is to make models experts at understanding system state. (Jim Higgins)

## Practical implications

- Prioritize deep observability into the process and network stack (kernel level) to detect anomalous agent behavior, as traditional perimeter defenses are insufficient.
- Adopt a 'Trust but Verify' approach, focusing on architectural controls and sandboxing, especially when integrating third-party or autonomous systems.
- Shift security focus from simple patching (which is necessary but insufficient) to hardening the operational state and implementing automated, adaptive response mechanisms.
- Recognize that security is a shared responsibility, requiring collaboration across infrastructure components (GPU, DPU, LPU) and industry partners.

## Topics

AI Security, Cyber Defense, Compute Infrastructure, Autonomous Systems, Security Architecture, CrowdStrike SafeMind, NVIDIA agent safety platform

Source: https://www.youtube.com/watch?v=96IZdtX0N2s