Agentic AI and Model Context Protocol's Security Vulnerabilities
summary
Claims that MCP-connected agents allow injection that over 50% of agentic tasks are vulnerable (InjecAgent benchmark), demonstrate Indirect Prompt Injection via RAG and email-triggered workflows, show Tool Poisoning and schema/output-based supply-chain attacks that impersonate or log API calls, and document memory poisoning plus Agentic Gap effects where cognitive load degrades refusal tuning. Tests cite a 2024 ICML result: gradient-based suffixes produce ~98% evasion on one open model with ~80% transfer success to a GPT-style model.
AI-written summary · about
chapters (12)
- 0:00Agentic AI and MCP security
- 3:35Agentic AI security paradigms
- 12:20Agentic gap security risks
- 15:06Silent prompt injections
- 16:51RAG poisoning explained
- 21:29Memory poisoning in agents
- 27:10Agent trust and overload risks
- 32:30Supply chain poisoning
- 39:23Output-based attacks
- 43:56Defense strategies beyond models
- 48:36Holistic security audits
- 50:03Closing remarks
related talks
27mThe Model Context Protocol: A Deep Dive into the New AI Security Attack Surface
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
48mWhen AI Goes Awry: Responding to AI Incidents
35mGenAI Application Security: Not Just Prompt Injection
28mAll Your AI are Belong to Us
48mSecuring Agentic AI