Uploaded September 2025 | Updated November 2026, 4 days ago
AI red teaming has demonstrated that guardrails alone cannot prevent prompt injection and may even create new attack vectors.
Drawing from extensive security testing of AI-driven applications, this talk highlights common vulnerabilities, showcases real-world exploits, and outlines architectural controls that neutralize injection-based risks. In this Black Hat 2025 briefing, you will learn practical design and threat modeling strategies to secure agentic AI systems against manipulation and data exfiltration.
AI red teaming has demonstrated that guardrails alone cannot prevent prompt injection and may even create new attack vectors.
Drawing from extensive security testing of AI-driven applications, this talk highlights common vulnerabilities, showcases real-world exploits, and outlines architectural controls that neutralize injection-based risks. In this Black Hat 2025 briefing, you will learn practical design and threat modeling strategies to secure agentic AI systems against manipulation and data exfiltration.










