In progress · Actively researching
Detection models & reconfigurable silicon
Compact models that inspect every request, and the longer job of compiling them onto reconfigurable chips so inspection stops costing what it costs today.
Inspection has to be cheap to exist
Anyone deploying agents inherits their attack surface, and the obvious defence — a model reading every prompt, retrieved document, tool call and output — adds latency and cost to every single request. At scale that bill is what decides whether the defence ships at all.
So the detection models have to be small: classifiers and embedding models that catch injected instructions, tool abuse and data exfiltration, built with the same compression toolchain as everything else here.
Down to hardware
Reconfigurable chips change the arithmetic. They run aggressively quantized models with streamlined logic rather than general-purpose floating-point hardware, and can be reprogrammed as the work changes — which is the property that makes fixed silicon viable for security at all.
This is where we are actively researching, and the roadmap sets out the phases it runs through.