Case Study: The Simulated Fleet
Three Agents. Ten Weeks. Every Shape of Losing It.
We built a simulated fleet of three AI personalities; a support agent, a coding agent and a research agent. And ran them through ten weeks of normal work, attacks, and drift. BRAIN_SCAN saw the fleet the whole time.
These are the measured results.
Two results deserve a closer look.
Precision. When BRAIN_SCAN paged for a prompt-injection burst, every single alarming slice contained genuinely hostile events. No crying wolf, the alarm meant something, every time.
Memory. When a surge in summarisation demand tripped the alarm, a human reviewed it, approved it, and the agent's library absorbed the new normal.
The same traffic re-scored clean, and the calibrated threshold did not move bit for bit. Alert fatigue is a design choice. We chose otherwise.
The Honest Limits
BRAIN_SCAN is built for the moment behavior changes shape. Very slow, sub-threshold creep surfaces as trend signals rather than alarms. We publish our limits because we measure them.
Full methodology and numbers available to design partners under NDA.