Vaikora › Blog › Product Updates
Eight milliseconds to block or allow: how Vaikora makes AI agent policy decisions before actions execute
The gap between an AI agent deciding to take an action and that action actually executing is the only window where enforcement is still cheap. Once the tool call completes, a file has been written, a record has been deleted, or an API has been called. Vaikora sits in that window and forces a deterministic decision on every action before it leaves the agent.
What "inline enforcement" means in practice
Most policy approaches work after the fact. A log gets written, an alert fires, a human investigates. That sequence is fine for detection. It is not enforcement, because the action already ran.
Vaikora sits as a proxy between the AI agent and whatever it calls: an LLM, an external API, a tool registered in an MCP server. Every request passes through the gateway. The gateway evaluates the request against policy and returns one of four decisions: allow, block, modify, or require-approval. The underlying system only receives the request if the decision is allow or modify. Block and require-approval stop execution at the gate.
The same input always produces the same decision. That determinism is deliberate. An enforcement layer that returns inconsistent results for identical inputs cannot be audited or trusted.
A concrete scenario
Imagine a customer support agent that has access to a CRM tool, an email-send tool, and an internal knowledge base. A prompt injection in a customer message attempts to redirect the agent to exfiltrate account records to an external endpoint.
Without a runtime gate, the agent calls the email-send tool with a payload containing account data. The call succeeds. The data is gone.
With Vaikora inline, the tool call hits the gateway first. Policy is configured to block any email-send action whose destination domain is not on the approved list and whose payload contains fields matching the account-data pattern. The gateway evaluates both conditions in the same pass, returns a block decision, and writes the full context to the audit log. The email never sends.
The enforcement latency at P50 is 8ms. At P99 it is 45ms. At 10,000+ actions per second throughput, this is not a bottleneck in any realistic agent deployment.
How policy gets defined and what gets checked
The gateway ships six content inspection modules. These cover things like data pattern matching, domain and URL reputation, payload structure, and tool identity. Policy rules combine these modules into conditions.
Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 give teams a starting baseline without writing rules from scratch. The preset defines which tool call categories are permitted, which data patterns trigger a block or require-approval, and which conditions escalate to a human reviewer before execution continues.
The gateway supports 12 LLM providers with no SDK migration. You point the agent at the Vaikora endpoint instead of the provider endpoint directly. The agent code does not change.
Audit chain and visibility
Every decision, allow, block, modify, or require-approval, gets written to a SHA-256 hash-chained audit log. Each entry references the previous entry's hash, so any tampering with historical records breaks the chain at that point. This is not just logging. It is a tamper-evident ordered record of every enforcement decision the platform made, which is what compliance teams and investigators actually need.
Vaikora integrates natively with Microsoft Sentinel, Microsoft Defender, CrowdStrike Falcon, SentinelOne, and AWS Security Hub. High-severity AI agent risk gets pushed into Azure Security Center and Microsoft Defender for Cloud as native incidents without custom engineering. Threats are pushed into SentinelOne as Threat Intelligence indicators with severity mapped numerically: Critical at 100, High at 75, Medium at 50, Low at 25.
Deployment path
The open-source component, Vaikora Guard MCP, is MIT-licensed, written in Python, published on PyPI as `vaikora-guard-mcp`, and self-hosted. It handles the inline proxy function and policy evaluation. The commercial control plane, hosted by Data443, adds centralized policy management, the compliance presets, the SIEM integrations, and the tamper-evident audit chain at scale.
It is also available as a SaaS subscription on AWS Marketplace with support for private offers and EDP draw-down, and as the Vaikora-AzureSecurityCenter solution in the Azure Sentinel content hub.
What to check next
If you are running AI agents today, the first question is whether any tool call currently executes without passing through a policy gate. Pull a sample of recent agent logs and look at what tools were called, with what payloads, and whether any of those calls would have been blocked under a minimal policy. That audit is free, takes less than an hour, and usually surfaces at least one category of call that should require approval before it runs.
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