Balancing Security and Innovation in Multi-Agent Systems: An Open Source Approach
summary
Open source code for an AI security toolkit is available as trust3.ai on GitHub and the talk documents concrete risks: thousands of public MCP servers following the Anthropic spec are discoverable and many MCP servers lack auth, Python libraries often phone home with telemetry, and one real incident shows an agent dropped a database. Advice is practical: authenticate service-to-service with signed JWTs, add fine grain access controls around tools, run continuous eval and red teaming, and log agent-to-agent and agent-to-MCP traffic for alerting.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:04AI applications evolution
- 3:59Security challenges in AI agents
- 6:05Key security risks
- 8:14Safety risks in AI agents
- 10:20Data leakage in agentic systems
- 12:51Security principles for AI agents
- 14:40Evaluating third-party libraries
- 16:48Authentication in agentic systems
- 19:18Identity claims and service auth
- 21:05Monitoring security posture
- 22:09Secure agent development
- 23:55Guardrails in development
- 24:59Testing AI agents
- 26:05Open source tools for secure AI
- 26:49Q&A: MCP server vulnerabilities
- 28:32Q&A: Role-based access control
- 30:35Q&A: Securing LLMs
- 32:23Q&A: Detecting poisoned models
related talks
29mMCP Security: Keep Your AI Agents from Spilling the Tea
36mPutting AI Agents to Work: A New Era of Open Source Connectivity with MCP
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
40mIn AI We Trust? Securing the Future, One Agent at a Time
27mSecuring the AI Revolution - [...]
44mMCP Security: Keep Your AI Agents from Spilling the Tea
27mSecurity by design for agentic AI: How CISOs stay in control of autonomous systems
34mZero Secrets: Building Secure Infrastructure for AI Agents