Vaikora › Blog › Product Updates
AI agent actions execute in milliseconds. Here is where Vaikora sits in that path and what it enforces.
Ten milliseconds is roughly the time between an AI agent deciding to call a tool and the tool executing. That window is where Vaikora operates. If a policy decision does not happen inside that window, it does not happen before the damage does.
This article explains the enforcement architecture: where Vaikora sits, what it checks, how decisions are made, and what the audit record looks like afterward.
Where Vaikora sits in the execution path
Vaikora installs as an inline proxy between the AI agent and whatever it calls. That includes LLM API calls, tool calls routed through the Model Context Protocol (MCP), and direct calls to external services. The agent never bypasses the gateway because the gateway is the path.
When the agent issues an action, Vaikora intercepts it, runs it through policy, and returns one of four verdicts before execution: allow, block, modify, or require approval. The action does not proceed until a verdict is issued.
This is the structural difference between enforcement and monitoring. A monitoring layer observes what happened. Vaikora decides what happens.
What the policy engine actually checks
Each intercepted action passes through content inspection modules. These check the content of the action against configured policy rules. A concrete example: an agent running inside a healthcare workflow attempts to include a patient record in an outbound API call. The content inspection module flags the payload as containing PHI. Vaikora's HIPAA compliance preset maps that finding to a block decision. The API call does not execute.
The same mechanism applies to prompts that attempt injection, to tool calls that exceed permitted scope, and to outputs routed to destinations outside policy.
Because Vaikora uses deterministic policy, identical input always produces an identical decision. There is no probabilistic scoring that lets the same payload through on Monday and blocks it on Tuesday. This matters for auditability and for consistency across high-volume agent workloads.
Vaikora reaches P50 enforcement latency of 8ms and P99 of 45ms, with throughput exceeding 10,000 actions per second. Those numbers mean the enforcement layer does not become the bottleneck in production agent deployments.
Compliance presets and what they configure
Out of the box, Vaikora ships presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. A preset is a starting policy configuration, not a certification. It maps the most common data types and prohibited actions for that regulatory context onto the enforcement rules.
A team deploying an agent that handles payment card data can load the PCI-DSS preset, review the rules, and begin enforcing on day one rather than writing policy from scratch.
Presets are a starting point. Organizations with specific risk profiles will extend them.
Audit records and the hash chain
Every decision Vaikora makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each record is signed into an ordered chain, so any attempt to delete or alter a record breaks the chain and is detectable.
For a compliance investigation or a security incident review, this means you can reconstruct exactly what an agent attempted, what verdict was issued, and when, in order, with integrity guarantees. That is the difference between an audit log and an audit record you can actually rely on.
How it connects to your existing security operations
Vaikora integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity AI agent threats surface in SentinelOne as Threat Intelligence indicators, with severity values mapped numerically so automated triage rules can act on them without manual configuration.
For Azure environments, high-severity AI risk promotes into Azure Security Center and Microsoft Defender for Cloud as native incidents. No custom engineering is required to get AI agent risk into the same queue where your analysts work.
Vaikora also adds a runtime enforcement gate on top of existing Azure RBAC, AWS IAM, and Snowflake roles. It does not replace those controls; it enforces an additional layer that those identity controls cannot see, because they operate at the resource level, not at the content and intent level of an agent action.
Deployment options
The Vaikora gateway ships as an open-source, MIT-licensed, self-hostable MCP-compatible proxy called Vaikora Guard MCP, published on PyPI. The enterprise control plane is hosted by Data443 and available on AWS Marketplace as a SaaS subscription with EDP draw-down support, and in the Azure Sentinel content hub.
It supports 12 LLM providers with no SDK migration, so existing agent code does not need to be rewritten to route through the gateway.
What to check next
If your organization is running AI agents in any workflow that touches regulated data, sensitive credentials, or external services, the first question to answer is whether any enforcement layer currently sits inline on those action calls. If your answer is a monitoring dashboard, that is detection, not enforcement, and it operates after execution, not before. Start by mapping which agent actions are currently uncontrolled, then review whether your compliance presets for the relevant regulatory context are configured and active.
See Vaikora enforce policy on your AI
Open-core AI runtime control. Self-host the MIT gateway free, or run the hosted Control Plane.
Get a demo Self-host the gateway
Vaikora