The Azure Security Podcast

Episode 111: Securing Agentic AI

10 snips
Apr 17, 2025
Amanda Minnich leads the Microsoft AI Red Team, focusing on the security of agentic AI systems. She dives into the evolution and capabilities of these advanced digital agents, emphasizing the expanded attack surface they create. The conversation highlights real vulnerabilities, like exploited chatbots, and stresses the importance of oversight. Best practices for securing AI agents are discussed, along with the unique challenges posed by large language models. Minnich also discusses the collaborative efforts needed to navigate the complexities of AI security.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Understanding Agentic AI

  • Agentic AI are digital entities that reason, plan, and act on behalf of users.
  • They connect to various tools and have memory, making them extremely powerful and versatile.
INSIGHT

Security Risks of Agentic AI

  • Agentic AIs can autonomously interact with complex data and act, increasing security risks.
  • Their expanded attack surface leads to risks like erroneous actions or successful jailbreaks.
ANECDOTE

Chatbot Car Sale Exploit

  • A car dealership chatbot was exploited with a simple jailbreak to authorize a $70,000 car sale for $1.
  • The chatbot's transactional power and no human oversight led to significant financial loss.
Get the Snipd Podcast app to discover more snips from this episode
Get the app