Why Security Matters: The Risks of Agentic AI and How to Mitigate Them
summary
Claims a practical permissioned access control that encapsulates Model Context Protocol (MCP) servers using Docker encapsulation and a policy enforcement engine, tested with the OpenAI agent SDK. It enforces file and domain-level read/write rules so prompt injections cannot access unauthorized paths, but cannot stop subtle parameter tampering (example: one-character change in a Bitcoin wallet). Implementation is compatible with existing MCP servers, imposes ~0.66 ms overhead, and successfully blocks many real-world misuse cases.
AI-written summary · about
chapters (8)
- 0:00AI security risks
- 1:03Rapid rise of agentic systems
- 3:57Agentic AI risks
- 6:09AI agent failures
- 7:53Mitigating risks with access control
- 11:05Testing safeguards
- 12:53Future behavior analysis
- 14:38Summary and discussion
related talks
29mMCP Security: Keep Your AI Agents from Spilling the Tea
51mAgentic AI and Model Context Protocol's Security Vulnerabilities
27mThe Model Context Protocol: A Deep Dive into the New AI Security Attack Surface
44mMCP Security: Keep Your AI Agents from Spilling the Tea
50mMCP at Scale: Governing AI Agent APIs When Autonomous Systems Are Your Consume…
22mSAFE-MCP: A Security Framework for AI+MCP (Model Context Protocol)
45mSecuring MCP Servers with Spring AI
27mSecuring the AI Revolution - [...]