VaikoraVaikora

Vaikora › Blog › Product Updates

Before the action runs: how Vaikora enforces policy on AI agents at the point of execution

Product Updates · Nov 20, 2026 · 4 min read

The gap in most AI security architectures is timing. Detection tools find problems after an agent has already written to a database, called an external API, or exfiltrated data through a tool response. By then the action is done and the audit trail is a post-mortem, not a control.

Vaikora addresses this by sitting inline between an AI agent and whatever it calls, whether that is an LLM, a tool, a database connector, or an external API. Every action passes through the enforcement gateway before it executes. The gateway returns one of four decisions: allow, block, modify, or require-approval. Nothing proceeds until a decision is made.

Why inline placement changes the security model

A monitoring layer tells you what happened. An inline enforcement layer decides what happens. The practical difference is that a blocked action produces no data exposure, no unauthorized write, and no lateral movement. There is nothing to investigate afterward because the event did not occur.

Vaikora applies deterministic policy enforcement, meaning identical input always produces an identical decision. That property matters for compliance teams: an auditor can replay any decision and verify the outcome. Every decision is written into a SHA-256 hash-chained, tamper-evident audit log, so the record cannot be altered after the fact.

Enforcement latency is P50 8ms and P99 45ms, with throughput exceeding 10,000 actions per second. Those numbers mean an agent running at production speed does not degrade because a policy gateway is in the path.

What the gateway actually inspects

The Vaikora gateway ships six content inspection modules. These cover the content of the action itself, not just metadata about which agent called which tool. A prompt injection attempt hidden inside a tool response, a PII string in an outbound payload, a request to write to a file path outside the declared scope: all of these are inspectable at enforcement time.

The gateway also supports 12 LLM providers without requiring an SDK migration. That matters for organizations running mixed environments where one team uses one provider and another team uses a different one. Policy applies consistently regardless of which model is behind the call.

Vaikora adds a runtime gate on top of Azure RBAC, AWS IAM, and Snowflake roles rather than replacing them. Identity permissions define what a principal is allowed to do in theory. Vaikora defines what an agent is allowed to do at the moment it tries to do it, with full context about the action content, the calling agent, and the active policy set.

A concrete scenario

Consider an AI agent built to query a customer database and generate reports. The agent has a database role that permits SELECT on several tables. Without a runtime gate, nothing stops the agent from being manipulated through prompt injection to SELECT from a table it should not touch, or to include PII in a response routed to an external endpoint.

With Vaikora inline, the policy set defines which tables are in scope, which data classes are permitted in outbound responses, and which external endpoints the agent may call. A SELECT against an out-of-scope table produces a block decision before the query executes. An attempt to route a response containing a PII match to an unlisted endpoint produces a block or a require-approval decision, depending on the policy. The agent never completes the action. The audit log records the attempted action, the policy that matched, and the decision, all signed into the hash chain.

Compliance presets and SIEM integration

For teams working toward specific regulatory requirements, Vaikora ships presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. A preset is a starting policy configuration, not a certification. Teams tune it against their actual environment.

For SOC visibility, Vaikora integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity decisions push into these platforms as native incidents or threat intelligence indicators, so an AI agent threat surfaces in the same workflow as any other security event.

What to check next

If your organization is running AI agents in production, the first thing to verify is where policy is currently enforced: at the identity layer, at the application layer, or at the point of execution. If the answer is "we review logs after the fact," the enforcement gap is real. Review which tool calls your agents are making today and whether any of those calls could produce an irreversible outcome if the agent were manipulated. That list is your starting policy scope.

The Vaikora gateway is open source under the MIT license and available as a self-hosted MCP-compatible proxy, so inspection of the enforcement logic does not require a commercial commitment.

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

More from the Vaikora blog