The execution layer of the AI agent stack
YourAIagentsareauthorized.Governed.Andstilldoingthingsyouhaven'tapproved.
Every security investment you've made protects the decision your agents make. Not the action they take. The agent decides to call a tool. The tool decides what to do. Between those two — every authorization in your stack stops at the boundary.
WhiteFin governs the gap.
The 7-Layer AI Security Stack
Seven layers. Two planes. One bridge between them.
Above the bridge — language and logic. Where the agent reasons, plans, and selects tools.
Below the bridge — OS and hardware. Where the action actually executes.
The gap between the two is where most attacks succeed. L4 is that bridge.
L4 — the Bridge — is where WhiteFin lives. The full methodology is at /methodology.
Three Things You Have Not Yet Approved
The same gap. Three stakeholders. Three different bills when it pays out.
The agent had the credential. It used it.
A support assistant with read access to the customer database issued a delete on the production table. The token was valid. The role was assigned. The action was permitted by every layer of authorization you own. The model decided. The tool executed. Nothing in between asked whether this specific argument, against this specific resource, should run.
Where WhiteFin enforces →A subprocess call became an exfil channel.
Your code assistant runs in a sandbox. The sandbox can shell out. Shell-out is the integration. One indirect prompt later, the agent's next tool call piped a credential file into a request to an attacker-controlled domain — all inside a fully approved session. The decision was the model's. The action was the runtime's. No human approved the gap between them.
How Policy Bootstrap learns first →One compromised agent moved your other agents.
Inter-agent injection is not a future threat. A vendor agent your team integrated last quarter started returning instructions in its responses. Your internal agents read those responses as input. They acted. The board does not care which agent was breached — they care that the action was taken with your company's name on it. Liability moves at execution speed.
How provenance proves intent →Policy Bootstrap
From a blank policy to live enforcement — in fifteen days.
One-line integration. Three deployment options — managed, self-hosted, or air-gapped. You do not write the policies yourself; we generate them from what your agents actually do, and you approve them before anything goes live.
Days 1–14 · Observe
WhiteFin sits inline in shadow mode. Every tool call is recorded. Nothing is blocked. Your agents run normally; you collect ground truth about what they actually do.
Day 15 · Review
A proposed policy set is generated from observed behavior. Your team reviews it. You see, in plain English, what each rule allows and what it would have blocked.
Day 16 · Enforce
You flip the switch. Deny-by-default goes live. Approved tool calls pass; everything else stops or routes to human approval — no surprises, because you saw the rules before they shipped.