Most of the difficulty with agents is not the model. It is that an agent sits between a user and a system that was designed on the assumption a user would be there in person.
This piece walks through how we think about it at Paycux, what we have changed our minds about, and where the sharp edges are.
What is prompt injection?
That brings us to what is prompt injection?. The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
An agent's credential should describe what it may do, not who owns it. Scope it to the narrowest set of operations that make the task possible, bind it to a single principal, and give it a lifetime measured in minutes rather than days.
Types of prompt injection attacks
Types of prompt injection attacks is where this gets concrete. Consent is the part most implementations get wrong. Asking once at install time and then acting indefinitely is not consent; it is a standing grant with no expiry and no visibility.
The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
- Scope every agent credential to one principal and one task
- Give tokens minutes of life, not days
- Record which agent acted, under whose authority, on what
- Make revocation a single call that takes effect immediately
Direct prompt injection
That brings us to direct prompt injection. The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
Consent is the part most implementations get wrong. Asking once at install time and then acting indefinitely is not consent; it is a standing grant with no expiry and no visibility.
Indirect prompt injection
That brings us to indirect prompt injection. An agent's credential should describe what it may do, not who owns it. Scope it to the narrowest set of operations that make the task possible, bind it to a single principal, and give it a lifetime measured in minutes rather than days.
The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
Consent is the part most implementations get wrong.
Multi-modal injection
Multi-modal injection is where this gets concrete. Consent is the part most implementations get wrong. Asking once at install time and then acting indefinitely is not consent; it is a standing grant with no expiry and no visibility.
The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
Agentic and tool-based injection
Agentic and tool-based injection deserves its own treatment. Consent is the part most implementations get wrong. Asking once at install time and then acting indefinitely is not consent; it is a standing grant with no expiry and no visibility.
The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
Notable real-world incidents
Consider notable real-world incidents. The audit trail matters more here than in any human flow. When something goes wrong, the question is never "did the user intend this" in the abstract — it is which agent, acting under whose authority, made which call, and whether the record can prove it.
Consent is the part most implementations get wrong. Asking once at install time and then acting indefinitely is not consent; it is a standing grant with no expiry and no visibility.
Where this leaves us
The pattern repeats across every system we have looked at: the hard part is not the mechanism, it is keeping the mechanism honest as the surrounding assumptions change.
If you are working through the same problem and want to compare notes, the docs cover the mechanics and the console shows the behaviour on your own data.
Everything here, already built
Sign-in, enterprise SSO, directory provisioning, roles and an audit trail behind one API. Start with the quickstart and have a working sign-in this afternoon.