Episode 32 · Season 1
Prompt Injection: 4 Practical Defenses for Your AI Agents
Agentic AI security is now the difference between “we shipped” and “we leaked.” In this episode, we break down the real security risks AI agents introduce, and the practical ways to mitigate them.
What you’ll learn:
- How prompt injection happens (and why LLMs struggle to separate trusted vs. untrusted instructions)
- How data exfiltration can occur even without a malicious prompt (often from system/design errors)
- How to reduce tool abuse risk by scoping tools, validating calls, logging, and using approvals for sensitive actions
Who it’s for: CTOs, CIOs, architects, engineers, and business leaders implementing or governing AI agents.
Guest + credibility: Greg is joined by Gavin Franklin and Tim to translate agent security into practical patterns teams can apply immediately.
CTA: Share this episode with someone responsible for AI rollout security. (And follow the show for more Agentic AI implementation guidance.)
Recommended Links
For questions or feedback, please contact us at CirriusTalk@CirriusSolutions.com
We cover
- Overview of AI Security Risks
- Deep Dive into Prompt Injection
- Understanding Data Exfiltration
- Exploring Tool Abuse
- Mitigation Strategies for AI Security Risks
Chapters
- 0:00Introduction to AI Security Concerns
- 2:51Overview of AI Security Risks
- 6:25Deep Dive into Prompt Injection
- 10:39Understanding Data Exfiltration
- 13:23Exploring Tool Abuse
- 15:16Mitigation Strategies for AI Security Risks
- 19:04Conclusion and Final Thoughts
Resources and downloads
The artifacts Greg’s team built for this episode, free to download.
Prompt Injection: 4 Practical Defenses for Your AI Agents
A walkthrough of three security risks that AI agents reintroduce: prompt injection, data exfiltration, and tool abuse. Each risk is traced through a worked email-agent example showing how it happens and what it costs, followed by the defenses that address them, including keeping secrets out of the model, human approval on sensitive actions, narrowly scoped sub-agents, and least privilege on every tool call.
Related episodes
Agentic AIBuild Your First Custom AI Agent: A Step-by-Step Walkthrough
Agentic AIAgentic AI Boundaries: Multi-Tenancy Done Right
Agentic AITop 3 Agentic AI Frameworks (and How to Choose Without Analysis Paralysis)
The Military MinuteFrom Military Spouse to Salesforce: Karri Webster’s PlaybookKarri Webster
Salesforce at ScaleUser Adoption Isn’t a Mandate, It’s a StrategyTalk to us about AI agents
Keep going: See how we build AI agents for real businesses, Talk to us about a custom AI agent, More from the Cirrius blog