Vaikora › Blog › Product Updates
AI agent took an action you didn't approve. Here is where the control should have been.
Every AI agent pipeline has a gap between "the agent decides to act" and "the action completes." In most deployments, nothing sits in that gap. The action executes, the log entry appears, and the investigation starts after the fact.
That sequence is the wrong order.
Why post-execution detection is not enforcement
Detection after execution means the file was written, the API was called, the record was modified. You can observe what happened. You cannot undo it reliably. In regulated environments, that creates a compliance problem. In production environments, it creates an operational one.
The correct control point is before execution. That means a policy decision must resolve before the agent's tool call or model output is forwarded. Allow, block, modify, or route to a human for approval. Those are the four outcomes. Anything else is just observation.
What Vaikora actually does, mechanically
Vaikora sits inline between an AI agent and the LLM or tools it calls. Every action passes through the enforcement gateway. The gateway evaluates the action against deterministic policy, issues one of the four decisions, and writes the result to a SHA-256 hash-chained audit log before anything downstream receives the request.
Deterministic means identical input produces identical output, every time. There is no probabilistic scoring that lets the same action through on Tuesday and blocks it on Wednesday. The policy is a set of explicit rules. The gateway applies them.
Enforcement latency is P50 8ms, P99 45ms, with throughput exceeding 10,000 actions per second. Those numbers matter because agents operate at machine speed. A gate that adds seconds of latency at scale breaks the pipeline. One that adds single-digit milliseconds does not.
A concrete scenario
Consider an AI agent connected to a customer database and an email API. A user prompt asks the agent to summarize recent support tickets and send a digest to the team.
Without a runtime gate: the agent queries the database (possibly retrieving more records than the prompt justified), formats the output, and calls the email API. If any of those records contain PCI-scoped card data, that data may leave the system before anyone reviews it.
With Vaikora inline: the database query action is checked against policy before it executes. If the query pattern matches a rule that flags broad retrieval from a PCI-scoped table, the action is blocked or modified to scope the query down. The email action is checked separately. If the content inspection module detects card data in the payload, the action is blocked. The agent receives a policy-enforcement response. The audit log shows both decisions, signed and ordered.
The compliance preset for PCI-DSS ships with Vaikora. No custom rule authoring is required to get baseline enforcement in place.
The audit log is part of the control, not an afterthought
Every policy decision Vaikora writes is signed into a SHA-256 hash chain. The chain is tamper-evident: any modification to a prior entry breaks the chain, making alteration detectable. This matters for two reasons.
First, it gives SOC analysts and compliance teams an ordered, verifiable record of every action an AI agent attempted and every decision the enforcement layer made. Second, it satisfies the auditability requirements that regulators and auditors look for when reviewing automated system controls.
An audit log that can be altered after the fact is not an audit log. It is a suggestion.
Integration with the security stack
Vaikora pushes AI agent threats into SentinelOne as Threat Intelligence indicators with severity mapping (Critical=100, High=75, Medium=50, Low=25). High-severity AI agent risk promotes into Azure Security Center and Microsoft Defender for Cloud as native incidents, with no custom engineering required. It also integrates with CrowdStrike Falcon and AWS Security Hub.
This means enforcement decisions made at the runtime gate appear in the platforms your SOC already uses. You are not building a parallel monitoring workflow.
What to check in your current setup
If you have AI agents in production or in a staging environment, ask these questions:
- What happens if an agent attempts to write to a resource it should not touch? Does something block the write, or does the log entry appear afterward?
- Can you produce a complete, ordered, tamper-evident record of every action an agent took and every decision a policy layer made?
- If an agent action is flagged as high-risk, does that signal reach your SIEM or EDR automatically, or does it require a manual export?
If any of those answers is "we check the logs afterward," the control is in the wrong place. The Vaikora gateway is open source, MIT-licensed, and published on PyPI as `vaikora-guard-mcp`. Start there, review the policy configuration, and test a block decision against a real tool call before you move further.
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