Vaikora › Blog › Product Updates
AI agent actions are not reversible: how Vaikora enforces policy before execution, not after
Reversing an AI agent action is often impossible. If an agent sends an email, posts to an API, deletes a record, or exfiltrates data to an external endpoint, the action is complete the moment the tool call returns. Detection after the fact tells you what happened. It does not undo it.
That is the core problem Vaikora is designed to solve.
Where the control point has to be
An AI agent operates by issuing tool calls: requests to read files, query databases, call external APIs, send messages, or modify records. Each tool call is a discrete, inspectable event with a known caller, a known target, and known parameters. That is exactly where policy enforcement belongs, before the call executes, not after telemetry arrives in a SIEM.
Vaikora sits inline between the agent and whatever it is calling. Every tool call passes through the enforcement layer. Vaikora evaluates the call against policy and returns one of four decisions: allow, block, modify, or require-approval. The tool call does not proceed until a decision is made.
At P50 the enforcement decision takes 8ms. At P99 it takes 45ms. For any workload running more than a handful of agents, Vaikora handles over 10,000 actions per second. The policy gate does not become the bottleneck.
What "deterministic" means in practice
Vaikora applies deterministic policy enforcement. Given identical input, the decision is always identical. There is no probabilistic scoring, no threshold to tune, no model that drifts. If a policy says block tool calls that target a domain outside an approved list, every call to an unlisted domain is blocked, every time.
This matters for auditability. Every decision Vaikora makes is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the log is tamper-evident and ordered. If a regulator or an internal investigation asks what an agent did at 14:32 on a given date, the answer is in the log and the log cannot be quietly edited.
Vaikora ships compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. A preset is a policy bundle, not a checkbox. It activates the specific controls those frameworks require for AI agent activity.
What the enforcement layer actually inspects
Vaikora ships six content inspection modules. These cover things like sensitive data patterns in tool call parameters, destination reputation, action type, and data classification. When an agent tries to pass a string that matches a credit card pattern to an external endpoint, the inspection layer catches it at the parameter level before the call leaves the system.
A concrete example: an AI agent is wired to a customer support tool chain. A user crafts a prompt that causes the agent to attempt a call to an external webhook with session tokens embedded in the query string. Without an inline enforcement layer, that call succeeds. With Vaikora, the call is evaluated against policy, the external destination fails the allow-list check, the tokens match a secrets pattern in the inspection module, and the call is blocked. The audit log records the blocked action, the matched policy rule, and the full call parameters.
The agent never knew it was stopped. The user never received confirmation the exfiltration worked.
How it integrates with existing security operations
Vaikora pushes blocked and high-severity decisions into the security stack rather than creating a parallel console. It integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.
For SentinelOne, Vaikora pushes AI agent threats as Threat Intelligence indicators with severity mapped numerically: Critical at 100, High at 75, Medium at 50, Low at 25. For Microsoft environments, high-severity AI agent risk is promoted into Azure Security Center and Microsoft Defender for Cloud as native incidents, with no custom engineering required.
SOC analysts see AI agent policy violations in the same queue as other alerts. They do not need a separate tool to investigate.
Vaikora also adds an enforcement gate on top of existing identity controls. It does not replace Azure RBAC, AWS IAM, or Snowflake roles. It adds a runtime layer that evaluates whether an agent with valid credentials should be allowed to take a specific action in a specific context, right now.
Where to start if you are evaluating this
The open-source component is Vaikora Guard MCP, published on PyPI as `vaikora-guard-mcp`, MIT licensed, written in Python, and self-hostable. It supports 12 LLM providers with no SDK migration required. You can wire it into an existing MCP tool chain and observe policy decisions against your own agent traffic before you configure enforcement.
Start by running in observe-only mode. Review what calls your agents are actually making, what parameters they are sending, and what destinations they are reaching. That inventory is usually the first surprise. Then build policy from what you see, not from assumptions about what your agents do.
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