Home | Blog | Eight Milliseconds to Block or Allow: How Vaikora Enforces AI Agent Policy Before Execution

Eight Milliseconds to Block or Allow: How Vaikora Enforces AI Agent Policy Before Execution

Giving an AI agent access to tools, APIs, and data stores without a policy gate is not a configuration oversight. It is a gap where the agent’s intended behavior and its actual runtime behavior can diverge, with no mechanism to catch the difference before damage is done. Vaikora sits inline between the agent and every resource it calls, and makes an allow, block, modify, or require-approval decision at that boundary before the action runs.

What "Inline" Actually Means

Most controls for AI agents are applied at the edges: prompt templates at the start, output filters at the end, or access controls at the infrastructure layer. These catch categories of risk but leave a gap during execution. The agent decides to call a tool, constructs a request, and that request reaches the target system before anything checks whether it should.

Vaikora works differently. The agent’s tool call passes through the Vaikora gateway. The gateway inspects the request, evaluates it against the active policy, and returns a decision. The tool call does not proceed until a decision is made. At P50 the decision completes in 8ms. At P99 it completes in 45ms. That is fast enough to sit in the critical path without changing how the agent or the tool is built.

The gateway ships six content inspection modules, supports 12 LLM providers, and handles MCP-compatible traffic. No SDK migration is required.

What Gets Checked

Policy evaluation at the gateway is deterministic. Identical input always produces an identical decision. This matters because it makes policy auditable and testable in a way that probabilistic approaches are not.

A concrete example: an agent working in a customer support context is configured with a policy that blocks any tool call that would write data to an external endpoint, requires human approval for any call that accesses records tagged as PII, and allows all other read operations within the workspace. When the agent constructs a request to export a customer record to a webhook, the gateway evaluates the request against that policy and blocks it. The agent receives the block response. The tool call never executes. The decision is written to the audit log.

The policy can be configured as a compliance preset or built from individual rules. Vaikora ships presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. The preset activates the relevant guardrails for that regulatory context without manual rule construction.

Audit Chain and Tamper Evidence

Every decision the gateway makes is signed into a SHA-256 hash-chained audit log. This gives an ordered, tamper-evident record of every action an agent attempted, what policy applied, and what the outcome was. The chain is not append-only in the trivial sense. Each record’s hash depends on the prior record, so any modification to historical entries breaks the chain at that point.

This matters for incident investigation. If an agent behaves unexpectedly and a security team needs to reconstruct what it did and when, the audit chain provides an ordered record that can be presented as evidence. It also satisfies the auditability requirements that show up in SOC 2 and ISO 27001 controls.

SIEM and EDR Integration

The gateway does not operate in isolation. Vaikora integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.

Specific threat signals get specific treatment:

  • High-severity AI agent threats are pushed into SentinelOne as Threat Intelligence indicators, severity-mapped at Critical=100, High=75, Medium=50, Low=25.
  • High-severity risk is promoted into Azure Security Center and Microsoft Defender for Cloud as native incidents, with no custom engineering required.

This means a blocked agent action that meets a severity threshold can generate an incident in the SOC’s existing tooling automatically. The analyst sees a Defender for Cloud incident, clicks through to the Vaikora audit entry, and has the full decision record including the policy that applied and the content that triggered it.

Deployment Options

Vaikora Guard MCP is the open-source component, MIT-licensed, written in Python, published on PyPI as `vaikora-guard-mcp`. It is self-hosted and MCP-compatible. Teams can run it in their own environment, evaluate policy behavior against their own agent traffic, and inspect the source.

The Vaikora Control Plane is the commercial product, hosted by Data443. It manages policy configuration, compliance presets, audit storage, and the SIEM integrations. It is available as a SaaS subscription on AWS Marketplace with support for private offers and EDP draw-down, and ships in the Azure Sentinel content hub as the Vaikora-AzureSecurityCenter solution.

Vaikora adds a runtime gate on top of Azure RBAC, AWS IAM, and Snowflake roles. It does not replace those controls. It adds a policy evaluation layer that those controls cannot apply because they do not inspect the content or context of an agent’s request, only the identity making it.

What to Check Next

If you are running agents against production data or external APIs, the first thing to verify is whether any control evaluates the content of a tool call before it executes. If the answer is only the agent’s system prompt, that is a gap. Install `vaikora-guard-mcp` from PyPI, route one agent’s traffic through it, and review the audit log after a session. The decisions the gateway makes and the ones it would have blocked are visible in the log before you configure any enforcement policy.