AI Agent Governance with Vaikora

AI agents that can act on production systems need governance that runs in real time, not after the fact. Vaikora evaluates every proposed agent action against deterministic policy before execution and produces a tamper-proof receipt for every decision.

Get a demo · Open-source gateway on GitHub · MCP server

The agent-action governance problem

An AI agent inside an application can read customer records, create invoices, send emails, post to Slack, run shell commands, and delete data. Each of those is a different risk profile. A read of a public document is fine. A delete of production records needs a human in the loop. A prompt with a leaked API key needs to be blocked outright. The control surface between the agent and the destination has to make those decisions in milliseconds.

Three-tier human-in-the-loop

Vaikora ships a pre-built CRUD policy matrix mapped to a three-tier approval model. Tier 1 reads auto-approve with silent logging. Tier 2 creates and updates require self-justification through the developer Slack OOB. Tier 3 deletes and merges escalate to the SecOps Slack with the agent held in a WAIT state until an admin signs off. Approvers click through to a web UI for the full action context and resolution token.

Built-in attack defenses

The Vaikora content modules detect prompt injection attempts, jailbreak patterns, leaked credentials, financial data exfiltration, and egress to known training-data endpoints. The defenses run synchronously inside the policy pipeline; bad actions never reach the LLM.

Audit and evidence

Every decision is signed into a SHA-256 append-only audit chain. Compliance presets cover SOC 2 Type II, HIPAA, GDPR, PCI DSS, and ISO 27001. Auditors can replay the chain and verify integrity without any vendor cooperation.

Related pages