Securing AI Agents...
summary
Claims threat modeling and demos, builds an intentionally vulnerable agent lab around a Notion-style orchestrator and subagents, and demonstrates bypassed prompt injection using mixed English/Tamil to escalate roles. Attack techniques shown include model and memory poisoning, authorization and control hijacking, role-inheritance and permission-chaining, and lateral agent-to-agent compromise; the demo reads CEO salary info after hijacking admin role. Defenses tested: least privilege, delegated OAuth-style tokens, expiring permissions, and real-time logging/alerts; OpenOrchestrator-style tooling is used for the lab.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI agents
- 1:05Agent fundamentals and operation
- 8:38Top 10 AI agent security threats
- 15:07Agent architecture deep dive
- 23:57Authorization and control hijacking
- 27:36Live hacking demo walkthrough
- 35:24Key takeaways and recommendations
- 36:28Q&A session
related talks
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
48mSecuring Agentic AI
27mKeynote: Towards Building Safe and Secure Agentic AI
53mHow to Break AI Systems (Before Someone Else Does)
19mKeynote: Towards Building Safe and Secure Agentic AI
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
33mSecuring Agentic AI With PyTorch: Threat Modeling & LLM Red Teaming in Practice
26mWhen AI Gets Manipulated: Security Risks in LLM Applications