Securing enterprise AI agent fleets.
The attack surface changed faster than the discipline defending it. We are building what comes next.
The problem nobody is defending
The tools we trusted were built for a world of human hands and static systems — a world that no longer exists.
AI agents now act. They read, decide, call, and execute — autonomously, at speed, across systems that were never designed to be touched by non-human actors.
The security tools we have were built for a different world. A world of enumerable identities, static endpoints, and human-readable intent. None of that applies here. The attack surface has changed faster than the discipline has.
We are working on what comes next.
Open research, before it is a product
We share the work in the open, while it is still research.
We publish detection rules, threat models, and security tooling for the agentic AI era — grounded in MITRE ATLAS, open source, and freely available. Everything we build is public before it becomes anything else.
Behavioural signatures for agent compromise, mapped to ATLAS tactics — usable the day they ship.
How autonomous agents are actually attacked: prompt-borne intrusion, tool abuse, and identity drift.
Instrumentation to see what agents do — and what is done to them — across the systems they touch.
अ + कवि — the unseeing seer
A name for the thing we are trying to build.
In the Vedic tradition, the Kavi perceives what ordinary sight cannot reach — the hidden pattern beneath surface appearance. The prefix a- does not negate. It elevates: beyond the limits of perception itself.
We named this organisation after the thing we are trying to build — a system that sees what AI agents cannot see about themselves.