Threat Modeling for Agentic AI: Stop Treating Agents Like APIs
Source: YouTube · Prabh Nair · published Aug 12, 2026 · 53:25
This video provides a practical framework for threat modeling agentic AI systems, emphasizing that security risks extend beyond traditional API protection to include prompt manipulation and agent behavior 0:10.
Key Takeaways:
• Agentic AI introduces unique threats like prompt injection and tool misuse, where agents execute unintended actions such as unauthorized refunds or data exfiltration 14:15
• Threat modeling must incorporate business logic and context, as attackers may exploit system permissions to cause financial loss by manipulating agent outputs 14:45
• A robust architecture includes classifier, responder, and QA reviewer agents, with human approval required for high-risk monetary transactions to prevent agent errors 10:30
• Security controls must address both technical vulnerabilities (e.g., identity spoofing) and human factors, such as analyst fatigue during manual approval processes 28:00
• Third-party integrations require rigorous assessment of their security frameworks (e.g., NIST, ISO) and clear liability terms to mitigate supply chain risks 31:00
• Threat modeling is a continuous process; documentation must be updated regularly to reflect new threats, architectural changes, and evolving regulatory landscapes 48:00
Effective security for agentic AI requires integrating traditional appsec principles with AI-specific frameworks like OWASP Top 10 for LLMs and MITRE ATLAS to identify and prioritize dynamic risks.
Sources:
Generate CPE Credits
Generate a professional CPE document from this video's transcript.
Estimated credit: 1 CPE hours
Estimate uses the video runtime (1 hour ≈ 1 CPE, rounded to the nearest 0.5, minimum 0.5, maximum 2.0). The final amount can be lower after review, never higher.
Topic: Cybersecurity. Commonly maps to: Security and Risk Management, Security Operations. Exact CISSP domains are assigned during generation.
Free account. One generation at a time, with a daily limit.
CPEBuddy is independent and not affiliated with or endorsed by ISC2, ISACA, or any certification body. Exports are formatted for common CPE submissions; acceptance is at your certification body's discretion.
Transcript Preview
First 800 characters of the transcript
What we going to cover today? >> So we are going to cover today about as a agentic [music] AI system and we'll be doing a threat modeling exercise for that system. >> What is agentic AI? >> In the traditional apps we had protected APIs. We just need to think about how the APIs were interacting. But with the agentic AI it is not only the APIs. It's like the promps and the agent how they are working. >> Whatever the name of the company and I'm booking ticket I want to put a cancellations cuz I have some other commitment. How it is a threat for the agentic systems. So first of all the biggest threat nowadays to the agentic system would be for this specific scenario is what if you are not the actual customer you are the attacker. >> Hi guys welcome to the session on coffee with prab and today …