AI agent security
How an agent proves what it is allowed to do, and how a counterparty checks that before acting on a request, so an agent cannot quietly overreach.
- Document
- Keynodex Agent-Trust Protocol (KATP)
- Status
- Working draft, not a published standard
- Type
- Research and protocol work in development
- Author
- Keynodex
Abstract
Abstract
The Keynodex Agent-Trust Protocol (KATP) is research and protocol work in development on AI agent security, agent-to-agent communication, and AI agent identity. It is working-draft research, not a published standard.
The Keynodex Agent-Trust Protocol, or KATP, is our name for this line of work. It is an ongoing research effort, not a finished specification. Nothing here should be treated as a stable contract or an integration target while the research is in progress.
The problem
Why agent trust is an open problem
As AI agents start to act for people, they run into a question the web never fully solved for software: how do you trust an actor you did not build? When an agent books, buys, replies, or negotiates on someone's behalf, the party on the other side needs to know who the agent represents, what it is authorized to do, and whether a human approved the action. Today those answers are ad hoc. Our research asks what it would take to make them verifiable.
Research areas
What we are researching
Agent-to-agent communication
How two agents acting for different people exchange requests and results with enough context to be trusted, without a shared owner in the middle.
AI agent identity
How an agent carries a verifiable identity tied to the person or organization it represents, so the other side knows who it is really talking to.
AI agent authentication
How authority and approval travel with a request, so an action taken on someone's behalf can be traced back to a human who signed off.
Doctrine
Human approval at the center
The through-line of this work is that a human stays in the loop for actions that matter. Trust rails are not just about machines vouching for machines. They are about making sure authority can be traced back to a person who approved it. That doctrine also shapes our products: it is why PingRep drafts and suggests rather than acting on its own.
An agent drafts a request
It says who it acts for and what it wants to do.
A person approves
Actions that matter wait for a human sign-off.
The action carries the approval
The other side can trace it back to the person who approved it.
Working draft
Working draft, not a published standard
KATP is early research and protocol work in development. We are not publishing a specification, making stability guarantees, or inviting integrations at this stage. If you are working on agent trust and want to compare notes, reach out.