Why our agents ask before they act

An automated defender has to be able to take action. It also has to avoid causing harm.

TOKEN/401 · Four minutes

Design principle · Response layer not built yet

An automated defender is only useful if it can take action: isolate a machine, revoke a credential, remove an agent’s access to a tool. It is only safe if it cannot take those actions carelessly. Containment that takes the payment system offline does the attacker’s job for them.

Our name comes from HTTP 401, the response a server sends when it needs to know who is asking before it continues. The same idea applies to an automated defender: before it changes anything important, it should check.

“Nothing the agents decide runs until it has been checked.”

In the response layer we are designing, every action starts as a proposal. Before it runs, it is checked for permission, likely impact, conflicts with other changes and whether it can be undone. This describes a system we are designing, not a released feature.

How much checking an action needs depends on what it could break. Rate limits, credential changes and isolating a machine can all affect legitimate users. Each organization should decide which actions can run automatically and which need a person to approve them.

After an action runs, the system should check its effect. A command finishing does not mean the incident is over. For now this is a design principle, while we continue to evaluate Tacit’s detection and how it behaves at runtime.

End of the essay