Скачать или смотреть Decoding AI Attacks with MITRE ATLAS

The landscape of AI risk is evolving at an accelerated rate, demanding a security framework built specifically for the unique attack surfaces of Machine Learning and Generative AI. Join host Pamela Gupta as she welcomes Walker Dimon, the MITRE ATLAS Lead, who is focused on advancing security for these rapidly evolving AI systems.
This conversation reveals the critical flow and severity of modern AI threats:
• Mapping the Adversary's Path: The MITRE ATLAS Matrix organizes the progression of attack tactics providing practitioners with a common language and taxonomy for AI threats.
• New, Realized Threats: The focus has shifted from predictive AI attacks (like data poisoning) to complex generative AI exploits. Walker explains that ATLAS techniques are only added if they are "realized"—meaning there is real-world evidence of actual adversaries using these TTPs against victim systems.
• The LLM Evolution: Learn about the need for new attacks taxonomies, including the recent addition of triggered injection, to capture the delayed adversarial behavior unique to complex Agentic AI systems.
The Value: Operationalizing Security, ATLAS provides the necessary blueprint for proactive defense:
• Walker explains how CISOs can immediately use ATLAS for threat modeling by mapping data flows and user access points to the matrix.
• It is a resource for mitigation strategies, offering strategies and exemplars like using open repository guardrail packages (e.g., Nemo guardrails) to define boundary conditions and prevent system compromise.
Tune in to understand the dynamic nature of AI risks and get actionable guidance on leveraging the MITRE ATLAS Matrix to build trustworthy, safe, and secure AI systems. We discuss Red Teaming, Prompt Injection attacks and a new category introduced "triggered injection".