Why detection and response need different systems

Attacks move in seconds, but deciding how to respond takes judgment. We think those two jobs need separate systems.

TOKEN/401 · Five minutes

Research · Tacit is in development

A stolen credential can be used seconds after it is taken, and an attacker can move across a cluster in minutes. Detection has to keep up with that. Ending an incident is different: you need to know where it started, what else was touched and what is safe to restore. Those answers come from investigation, and investigation takes time.

So we are building the system in two parts. Tacit, a streaming model, watches events continuously. A team of AI agents would then investigate what Tacit flags and plan a response. Our work today is on Tacit. The agent layer is still a research direction.

“Detection has to be fast. Deciding what to do about it has to be careful.”

Tacit reads events and the relationships between processes, credentials and resources, and keeps that context as the stream continues. We evaluate it on what it remembers and where it gets things wrong. A good score on a side task does not make it a reliable risk score, and we do not treat it as one.

In the agent layer, an agent would examine the evidence, test possible explanations and weigh the impact of a response. Investigation, containment, recovery and verification would stay separate jobs.

The open question is the handoff between the two parts: what context Tacit should pass on, which decisions should wait, and how what the agents learn could improve Tacit without interrupting it.

We have not yet measured detection accuracy, false positives in unfamiliar environments or behavior under sustained load. This article describes where the work is going, not a finished product.

End of the essay