Your agents see everything. Today, you cannot prove what they did with it.
AI agents move faster than your existing controls. They read across systems your IAM was never designed for. They make decisions your SIEM cannot see. And when a regulator, an auditor, or your own board asks what happened and why, the answer lives in model logs that nobody outside engineering can interpret.
For most organisations the response so far has been to keep agents away from anything sensitive. That gets harder every quarter. Vortalis exists so you can deploy agents into the work that matters, with controls your security, compliance, and risk teams can stand behind.
Three jobs. One trust layer.
Tokenise.
On actions routed through the service proxy, the fields an adapter declares as sensitive are replaced with safe substitutes before they reach the agent, so it works on stand-ins scoped to the job, not on the raw values. The original values stay inside your trust boundary, exposed only when policy allows it and only with an auditable reason. Coverage is per adapter action and per declared field; the MCP gateway and the governance contract do not tokenise.
Enforce.
Your policy decides what the agent is allowed to do, not the model. Out-of-policy actions routed through the service proxy are blocked or routed for human approval before they reach a system of record. MCP tool-call hooks and the LLM ingress are opt-in or EXPERIMENTAL, so this does not yet cover every channel an agent might use.
Audit.
Each decision the trust layer makes about an action routed through the service proxy is written to a per-tenant, SHA-256 hash-chained, tamper-evident log. Audit rows are chained, not individually signed; exports and evidence packs are signed. Your audit team or external auditor can verify an export offline. They never need to call our servers, and they never depend on us being available.
Six stages, working together.
- 1
Kill Switch.
Checked first, before any other stage runs. Pause one agent, one service, a tenant, or every agent on the platform. Other running instances refresh about every five seconds, so a change is not instant everywhere.
- 2
Policy Evaluation.
Your rules, including human-in-the-loop approval for high-impact actions, applied at runtime to every action routed through the service proxy; a soft deny creates an approval request. Soft-deny release is built and off by default: when enabled, an approver's decision releases that exact request once, within 24 hours of approval; service proxy only; intent and joint-action holds only. MCP tool calls and the LLM ingress are covered only where those opt-in paths are switched on.
- 3
Service Execution.
The permitted action runs against the downstream service. Custom-code adapters are refused on the production platform today, because the host cannot provide the confinement the EXPERIMENTAL sandbox needs.
- 4
Data Protection.
Declared sensitive fields on service-proxy actions are tokenised before entering agent context. Exposure of original values is gated by policy.
- 5
Audit Logging.
Hash-chained records of every decision on service-proxy actions, with signed exports. Verifiable offline.
- 6
Anomaly Detection.
Tenant-scoped statistical baselines flag unusual patterns on service-proxy traffic. A critical data-access anomaly can engage the kill switch automatically; other detections raise alerts.
Built for regulated industries from day one.
Vortalis maps to the major UK and EU regulatory frameworks for AI in financial services and other regulated sectors, alongside the leading AI-specific and information-security standards. The mapping is current, maintained, and produced as evidence in a format your auditor recognises.
Get the regulatory coverage mapThe sectors Vortalis is built for.
- Banking and capital markets
- Insurance
- Asset and wealth management
- Healthcare
Built for the people who own the risk.
The role selector at the top of this page tailors the headline. These pages go deeper for the three roles who own the risk day to day.
Talk to us before you deploy your next agent.
Send us your use case and we will share a signed evidence pack drawn from a real agent action analogous to your scenario. No slide deck. No NDA required for the first exchange.
See a signed evidence pack