An imagined Manhattan bank vault stands open, with shelves of cash visible inside

We rebuilt the bank breach and reattacked the fix.

We recreated a bank permission flaw in Northline, our working bank with fictional customers. An AI agent reached private records. The report shows exactly how they were exposed, then what happened when we fixed the flaw and attacked again: zero private records exposed.

Read the investigation
What we research

Intent has to survive the trip.

We follow agent actions from the instruction a person gave to the effect a system recorded. Each research note shows the control question, the evidence, and what the result means for teams shipping agents.

The control layer between intent and action

Four maintainer-reviewed fixes carry human control through cancellation, desktop retries, scheduled recovery, and reviewed saves in CopilotKit’s OpenDots and OpenMuse.

Agent securityExecution integrityHuman control
Read research note

Frontier AI Programs

Work with Kronowave

Your agent ran. What did it actually do?

Bring us the workflow that matters to your next release or customer. We trace the action to the system that received it, show what happened, and help your engineers verify the change.

48-hour Rapid Scan