Runtime control for AI agents. Not after-the-fact log analysis.

Your IAM was not designed for agents. Your SIEM cannot see their reasoning. Vortalis is the trust layer for AI agents: it decides what agents are allowed to do at runtime, before they do it, and produces a hash-chained record, with signed exports, that your SOC can actually consume.

Three problems your existing stack does not solve.

1

Over-broad agent access.

Agents inherit the data scope of the human or service account they impersonate. That scope was set in a world without agents. On actions routed through the service proxy, Vortalis tokenises the fields an adapter declares sensitive before they reach the agent, so the agent operates on safe substitutes and does not hold those raw values. Coverage is per adapter action and declared field; the MCP gateway and the governance contract do not tokenise.

2

No runtime guardrail.

Prompt-side guardrails are advisory. Vortalis applies your policy on the action path. Out-of-policy calls routed through the service proxy are blocked or sent for approval before they reach a system of record. The model is not the enforcement point, so there is no model-side bypass on that path; MCP tool-call hooks and the LLM ingress are opt-in or EXPERIMENTAL.

3

Decisions your SIEM cannot see.

Model logs are not security telemetry. Vortalis writes a hash-chained event for each decision on a service-proxy action, and signs exports, in a format your SIEM, your SOAR, and your audit team can all consume. One source of truth, not three.

Designed to sit alongside the controls you already run.

Vortalis is a control plane between your agents and your data, not a replacement for your IAM, your DLP, your SIEM, or your model gateway. We integrate with the systems you already pay for, and we do not ask you to rip anything out. Your existing identity, network, and data-loss controls continue to do their job. Vortalis adds the layer that did not exist: runtime decision control for non-deterministic actors.

What you get.

  • A runtime decision point that evaluates every action routed through the service proxy before it executes.
  • Tokenisation of declared sensitive fields on service-proxy actions, with policy-gated exposure of originals.
  • A kill switch that pauses one agent, one service, a tenant, or the entire platform, with other instances refreshing about every five seconds.
  • Hash-chained audit records with signed, offline-verifiable exports, suitable for SOC, audit, and incident response.
  • Integration paths that exist today: 29 service adapters (adapter shapes, not counterparty-certified), Python and TypeScript SDKs, and an OpenAI-compatible LLM gateway. The LLM ingress is EXPERIMENTAL and off by default.
  • Governed inter-agent delegation: every action taken under an explicit delegation grant is checked against policy before it runs, with the full chain recorded in the hash-chained audit trail. Delegation and joint-action checks are built as options, off by default and enabled per tenant. Where joint-action evaluation is on, a rule can also match a syntactic, time-windowed pattern over audited event text. There is no semantic understanding of what agents say to each other, and this does not solve emergent multi-agent coordination in the general case.
  • Intent-vs-action drift telemetry that measures the gap between a session's declared intent and its actual action stream, alerting before it becomes a violation. This is not general goal-drift detection and is not a substitute for model-alignment research.

A governance proxy has a real defensive perimeter and a real outside of that perimeter. Before you bring Vortalis into a procurement conversation, read what we do not defend against.

What Vortalis does not defend against

Bring your architect to the next conversation.

Send us a real agent action you are evaluating. We will return a decision record, exported in a signed evidence pack, and the policy that produced it, on your stack pattern, ahead of any meeting.

See a signed evidence pack