Agentic AI Security: How Microsoft Prevents Autonomous Agent Attacks?
As agentic AI systems move into the mainstream—powered by tool calling, MCP, and autonomous workflows—security is no longer a “nice to have.” It’s mission-critical.
In this episode, we sit down with Raji, Principal Engineer & Manager for AI and Safety at Microsoft, to deep-dive into the rapidly evolving world of AI security, autonomous agents, and enterprise governance.
Discover how Microsoft identifies and mitigates risks in agentic AI, distinguishes AI Security vs AI Safety, and enables organizations to deploy autonomous systems safely at scale—without slowing innovation.
Key Insights & Moments:
01:05 – Meet Raji: Cybersecurity leader and Principal Manager for AI Safety at Microsoft. 🚀
05:15 – OWASP Top 10 for Agentic AI📋
07:15 – The Model Context Protocol (MCP) 🌐
08:15 – AI Security vs. AI Safety ⚖️
09:40 – Socio-Technical Systems🌍
18:15 – Three Buckets of Testing🧪
23:35 – Governing Agents at Scale📏
25:20 – The "Brakes" Analogy 🏎️
27:10 – Documentation and Context📝
🎯 Key Topics Covered
🔹 AI Security vs AI Safety — what’s the real difference?
🔹 OWASP AI Security Top 10 for autonomous agents
🔹 Prompt Injection attacks & model-level vulnerabilities
🔹 MCP (Model Context Protocol) & tool-calling security risks
🔹 Microsoft’s 3-Bucket Testing Framework
• Application
• Model
• Output
🔹 Enterprise governance for scaling agentic AI responsibly
🔹 Real-world case studies: Healthcare, Finance & Loan Bias
🚨 Why This Matters
Agentic AI systems take autonomous actions without human intervention. Unlike traditional applications, they operate in unpredictable ways—making them vulnerable to inference manipulation, data leakage, and misuse. Raji shares Microsoft's battle-tested approach.
👥 Perfect For
✅ AI security researchers & practitioners
✅ Enterprise architects deploying autonomous agents
✅ Developers building tool-calling workflows
✅ CISOs & compliance leaders
✅ AI engineers scaling models responsibly
🚀 Building AI You Can Trust
As organizations adopt autonomous AI systems, security and privacy must scale with innovation. @protectoai enables enterprises to safely deploy AI by protecting sensitive data at every layer.
🌐 Learn more: https://www.protecto.ai/
🔗 Follow on LinkedIn: https://www.linkedin.com/company/protectoai/
🔔 Subscribe for in-depth discussions on AI security, governance, and responsible AI: https://www.youtube.com/@protectoai/
#AgenticAI #AISecurity #AgenticAISecurity #MicrosoftAI #AutonomousAgents #DataSecurity #EnterpriseAI #AIGovernance #Cybersecurity #SecurityResearch #OWASP #BlackHat #AIVulnerabilities #ProtectoAI