A platform we are actively building — a runtime governance layer that understands enterprise AI as a graph of users, agents, models, tools and data, and controls what each can actually reach. This is our forward view and current R&D focus, not a completed client engagement.
A single control layer for security, policy, risk, observability and compliance — across every model and agent.
Every department is shipping copilots and agents. Each one authenticates differently, logs differently, and reaches into real systems — CRM, ERP, payments, HR. Security and compliance teams are asked to sign off on an estate they can't see, using tools that only inspect the text of a prompt.
Autonomous agents now call tools and APIs on their own. User-level RBAC never anticipated a non-human identity that acts across systems.
A prompt looks harmless in isolation. What matters is what it can reach — payroll, PII, an external model — and today that's rarely mapped.
When something goes wrong, teams can't reconstruct who asked what, through which agent, to which system. There's no graph to query.
Borrowing attack-path analysis from cybersecurity: follow a single request past the model and you see what it can actually touch. The question governance must answer isn't "is this text allowed?" — it's "what is the blast radius?"
Blast radius: this request can reach customer PII and push it to an external service. A keyword filter waves it through — a graph that knows the path stops it, or routes it to human approval. Illustrative path.
Users, sessions, agents, policies, models, tools, APIs and data assets become nodes; every request is a path through them. Governance decisions are then made on reach and permission — not the words in a prompt.
An autonomous agent is a new kind of identity — it acts on its own. It needs fine-grained, per-action permissions, scoped to exactly what its job requires and nothing more. The graph makes those permissions enforceable, because it knows what each action would actually reach.
Illustrative policy for a single agent.
A single "risk: high" tells an executive nothing actionable. We score each dimension separately, with a plain-language reason — so the team knows why a system is risky and what to do about it.
Illustrative — one AI applicationIllustrative risk exposure by dimension (higher = more exposure).
This is an active initiative. We'd rather build it in the open with a handful of enterprises than behind closed doors.
Discover the AI estate — apps, agents, models, tools, data — and build the relationship graph. Read-only, in your environment.
Route AI traffic through the runtime to observe real paths, then produce risk intelligence and trust scores grounded in what actually happens.
Turn on the policy engine and agent permissions — governance-as-code, with human-approval routing on high-risk paths.
Combined in a way that's designed and implemented independently. See the full platform overview →
We're looking for enterprises deploying AI at scale to build and validate this with. Let's talk about your governance gaps.
