@ericyocam6Professor, California Polytechnic State University
How Are You Securing Your Agentic AI Systems?
Which tool or approach are you using to secure your Agentic AI systems against prompt injection and unauthorized actions?
@ericyocam6 @ericyocam6
Comments3
For me it's mostly least-privilege: agents only get the tools they need, plus my approval on anything that sends, deletes or spends money. I'd vote for that option.
Share the final vote counts when this closes, plus the option you'd pick yourself. Founders selling agents to bigger companies could use that list when a buyer asks how they stop an agent from being tricked into taking actions nobody approved.
Thanks, Miles. I agree. Least-privilege access is one of the most effective ways to reduce the risks associated with autonomous agents. Requiring human approval for sensitive actions, such as sending information, deleting data, or making financial transactions, adds another important layer of protection. My research also focuses on data provenance, ensuring that untrusted information cannot influence an agent's actions simply because it came through another agent or tool. I think combining these approaches provides a stronger security foundation. Appreciate you sharing your approach!
Thanks, Ryan. I'll share the final poll results once voting closes. My preference is a layered security approach combining least-privilege access, data provenance verification, and human approval for high-impact actions. My recent research on steganographic prompt injection demonstrated that traditional content filtering alone isn't sufficient, particularly when malicious instructions are concealed in images or metadata and passed between agents. Provenance-based enforcement helps address that risk by tracking where information originated and preventing untrusted content from triggering unauthorized actions. I agree with Miles that restricting tool permissions is an essential part of the solution. Ultimately, securing agentic AI requires multiple complementary controls rather than relying on a single defense.
3 comments
For me it's mostly least-privilege: agents only get the tools they need, plus my approval on anything that sends, deletes or spends money. I'd vote for that option.
Share the final vote counts when this closes, plus the option you'd pick yourself. Founders selling agents to bigger companies could use that list when a buyer asks how they stop an agent from being tricked into taking actions nobody approved.
Thanks, Miles. I agree. Least-privilege access is one of the most effective ways to reduce the risks associated with autonomous agents. Requiring human approval for sensitive actions, such as sending information, deleting data, or making financial transactions, adds another important layer of protection. My research also focuses on data provenance, ensuring that untrusted information cannot influence an agent's actions simply because it came through another agent or tool. I think combining these approaches provides a stronger security foundation. Appreciate you sharing your approach!
Thanks, Ryan. I'll share the final poll results once voting closes. My preference is a layered security approach combining least-privilege access, data provenance verification, and human approval for high-impact actions. My recent research on steganographic prompt injection demonstrated that traditional content filtering alone isn't sufficient, particularly when malicious instructions are concealed in images or metadata and passed between agents. Provenance-based enforcement helps address that risk by tracking where information originated and preventing untrusted content from triggering unauthorized actions. I agree with Miles that restricting tool permissions is an essential part of the solution. Ultimately, securing agentic AI requires multiple complementary controls rather than relying on a single defense.