Vaikora › Blog › Product Updates

An AI agent about to run a shell command: what stops it, and how does the decision get made?

Product Updates · Nov 11, 2026 · 4 min read

An AI agent can call a tool, write a file, query a database, or run a shell command without pausing unless something sits between the request and execution. Without a policy gate, the only controls are the permissions already granted to the underlying identity. That is usually too broad.

This article explains how Vaikora's runtime enforcement works mechanically: where it sits, what it checks, and how decisions are recorded.

Where the gate sits

Vaikora places itself inline between an AI agent and whatever the agent is trying to call. That could be an LLM API, an MCP tool server, a database endpoint, or an external service. Every action the agent attempts passes through the Vaikora gateway before execution.

The gateway ships six content inspection modules. When an action arrives, the relevant modules evaluate it against the policy set configured for that agent context. The result is one of four outcomes: allow, block, modify, or require-approval. The action does not proceed until a decision is returned.

That decision is deterministic. Identical input always produces identical output. There is no probabilistic scoring that lets a borderline action through on a good day and blocks it on a bad one.

What the latency looks like in practice

Speed matters for inline enforcement. A gate that adds seconds of overhead per action breaks agentic workflows.

Vaikora runs at P50 8ms and P99 45ms. At that latency, a multi-step agent workflow calling ten tools in sequence adds under half a second of enforcement overhead in the worst-case percentile. The gateway handles 10,000+ actions per second, so throughput is not a constraint in most deployments.

Consider a concrete example. An AI coding agent is asked to push a configuration change to a production database. It constructs a write query and calls the tool. The Vaikora gateway intercepts the call, checks the query content against policy (data classification scope, environment label, approval requirements), and evaluates the agent's identity context. If the policy says production writes require human approval, the action is held and an approval request is dispatched. The write does not execute until the approval arrives or the timeout triggers a block. The agent receives a structured response explaining the hold.

How the audit trail works

Every decision Vaikora makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry in the chain includes the prior hash, so any modification to a past record breaks the chain and is detectable.

This matters for compliance investigations. If a regulator asks what an AI agent did with a specific dataset on a specific date, the audit log provides an ordered, verifiable record of every policy decision made against that agent's actions. The log does not depend on what the agent reported about itself.

Vaikora ships compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. Each preset defines the policy scope and logging requirements that apply to those frameworks without requiring teams to build policy sets from scratch.

How it integrates with existing security tooling

Vaikora does not replace RBAC, IAM, or Snowflake roles. It adds an enforcement layer on top of them. An action that is permissible under IAM can still be blocked by Vaikora policy if it violates content rules, approval requirements, or scope constraints.

For teams already running Microsoft Sentinel, SentinelOne, CrowdStrike Falcon, or AWS Security Hub, Vaikora integrates natively. High-severity AI agent risk surfaces in Microsoft Defender for Cloud as native incidents without custom engineering. Threats push into SentinelOne as Threat Intelligence indicators with severity mapped numerically: Critical at 100, High at 75, Medium at 50, Low at 25.

Vaikora supports 12 LLM providers with no SDK migration required. The open-source gateway (vaikora-guard-mcp, MIT-licensed, available on PyPI) handles MCP-compatible traffic and can be self-hosted. The commercial control plane, hosted by Data443, handles policy management, audit log storage, and integration with the broader security stack.

What to check next

If you are evaluating whether your current AI agent deployment has an enforcement gap, start by mapping what tools each agent can call and whether any of those calls touch production data, credentials, or external services. If the only control is the IAM role attached to the agent identity, there is no content-level policy, no approval path, and no tamper-evident record of what executed.

Pull the vaikora-guard-mcp package from PyPI and run it against a test agent workflow to see what the policy decision log looks like for your actual traffic patterns. The enforcement latency numbers become concrete quickly once you have real tool calls going through the gate.

More from the Vaikora blog