Vaikora › Blog › Product Updates
What actually runs between your AI agent and the tools it calls, and what can stop it
An AI agent can read a database, write a file, call an API, or delete a record. The only question that matters for security is whether anything checks that action before it executes, not after.
Most agent architectures today have no gate at that point. The agent gets a task, the LLM picks a tool, the tool runs. If the action was out of policy, you find out from a log review or an incident.
Where the control point has to sit
The window between "tool call requested" and "tool call executed" is where enforcement belongs. Once a tool call completes, the action has already happened. You can detect it, record it, alert on it. You cannot reverse most of it.
Vaikora sits inline at that exact point. Every action an AI agent wants to take passes through the Vaikora gateway before it reaches the LLM or the tool. The gateway applies a deterministic policy decision: allow, block, modify, or hold for human approval. The decision runs at P50 8ms and P99 45ms, so the latency added to any agent workflow is real but small.
Deterministic means what it sounds like. Identical input produces identical output every time. There is no probabilistic scoring, no "usually blocks this category." The policy rule either matches or it does not.
What the gateway actually inspects
Vaikora ships six content inspection modules. At runtime, each action is checked against the active policy set before execution continues. The kinds of things a policy rule can target include:
- Tool name and parameters (block any call to `delete_record` where the target table is in a protected schema)
- Data in the request or response payload (block exfiltration of anything matching a PII pattern)
- Agent identity and context (require approval when an agent operating in a production environment requests a write operation outside business hours)
- Compliance preset violations (HIPAA, PCI-DSS, GDPR, SOC 2, ISO 27001 presets ship out of the box)
A concrete example: an agent is tasked with summarizing customer support tickets. Its system prompt authorizes read access to the ticketing system. Mid-session, a prompt injection in one ticket instructs the agent to call an outbound webhook with the full ticket dataset. Without a gate, that call goes out. With Vaikora inline, the outbound webhook tool call hits the policy engine. If no rule permits outbound webhook calls for that agent identity, the call is blocked before it executes, and the decision is written to the audit log.
Auditability that holds up under scrutiny
Every decision Vaikora makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry is signed into an ordered chain, so any modification to a past record breaks the chain at that point. For compliance investigations, that means you can demonstrate exactly which policy rule applied to which action at which timestamp, with evidence that the record has not been altered.
That matters for PCI-DSS scope discussions, HIPAA breach analysis, and SOC 2 audits. The audit trail is the enforcement record, not a summary of what might have happened.
How this fits into existing security infrastructure
Vaikora integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity AI agent threats push into SentinelOne as Threat Intelligence indicators with severity mapped numerically (Critical = 100, High = 75, Medium = 50, Low = 25). High-severity risk promotes into Azure Security Center and Microsoft Defender for Cloud as native incidents, with no custom engineering.
This means AI agent enforcement signals appear in the same workflow your SOC already uses for other threat types. Analysts do not need a separate console to investigate an agent policy violation.
Vaikora also adds its runtime gate on top of Azure RBAC, AWS IAM, and Snowflake roles rather than replacing them. If IAM says an agent identity can access a resource, Vaikora can still block or modify that access based on the action context. The two layers answer different questions.
What to check in your own environment
Before evaluating any AI agent security tooling, map the execution path of one agent in your environment and answer these three questions:
- At what point in that path does a policy decision currently get made?
- If the agent is injected mid-session, which actions could it take before detection?
- Does your audit log record the specific tool call that was made and what data it touched, with evidence of log integrity?
If the answers are "after execution," "several," and "no," the enforcement gap is real. The place to start is the gateway layer, not the monitoring layer.
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