@ericyocam6Professor, California Polytechnic State University
Can AI Agents Trust Other AI Agents?
Which approach are you using to prevent malicious or untrusted data from influencing AI agent decisions?
@ericyocam6 @ericyocam6 @ericyocam6Which approach are you using to prevent malicious or untrusted data from influencing AI agent decisions?
Comments2
Your title asks whether one agent can trust what another agent hands it, so share one example from your own work where bad information got passed along and got caught before it changed a decision. A founder whose buyer asks what happens when one agent feeds another something wrong could point to that story.
Thanks, Ryan. A good example comes from my recent research on steganographic prompt injection in agentic AI systems. We demonstrated how malicious instructions could be concealed within seemingly harmless images, such as an email signature logo, and then extracted by a tool and passed to another agent as if they were trusted instructions. In our experiments, this created a path for unauthorized actions. We developed a defense using decode-time provenance binding and HMAC-signed trust labels to track the origin of extracted content and block unauthorized tool calls. The key takeaway is that an agent shouldn't automatically trust information simply because it came from another agent or tool. Trust needs to follow the data throughout the workflow. Appreciate the suggestion!
2 comments
Your title asks whether one agent can trust what another agent hands it, so share one example from your own work where bad information got passed along and got caught before it changed a decision. A founder whose buyer asks what happens when one agent feeds another something wrong could point to that story.
Thanks, Ryan. A good example comes from my recent research on steganographic prompt injection in agentic AI systems. We demonstrated how malicious instructions could be concealed within seemingly harmless images, such as an email signature logo, and then extracted by a tool and passed to another agent as if they were trusted instructions. In our experiments, this created a path for unauthorized actions. We developed a defense using decode-time provenance binding and HMAC-signed trust labels to track the origin of extracted content and block unauthorized tool calls. The key takeaway is that an agent shouldn't automatically trust information simply because it came from another agent or tool. Trust needs to follow the data throughout the workflow. Appreciate the suggestion!