AI governance isn't a dashboard problem — it's an infrastructure problem. Kakunin is the API layer that enforces governance at the machine level: cryptographic agent identity, continuous behavioral monitoring, automated revocation, and immutable audit trails. Built for MiCA and EU AI Act compliance from day one.
Most AI governance tools sit above your stack. They watch. They visualise. They alert. Kakunin sits in your stack — in the request path, at the gateway layer, in the event ingestion pipeline. Governance controls are enforced in real time, not reviewed after the fact.
A gateway that calls GET /api/v1/verify/{serial} on every inbound agent request does not just log a compliance event — it blocks a non-compliant agent before any business logic runs. Kakunin governance controls are operational controls, not reporting overlays.
When an agent's behavioral risk score crosses 0.85, the revocation happens in under 60 seconds with no human in the loop. The governance action — withdrawing the agent's operating authority — occurs at machine speed, not at the pace of a compliance review cycle.
Kakunin integrates into your existing stack via REST API and TypeScript SDK. No separate governance portal for operators to maintain. No parallel data pipeline to keep in sync. The governance layer is the API layer — the same endpoints your agents call are the endpoints your compliance team queries.
Identity records, behavioral events, risk scores, revocation history, and compliance reports all live in one tamper-proof system. When a regulator asks for audit evidence, the answer is a single API call — not a spreadsheet assembled from five different monitoring tools.
Governance without continuous monitoring is compliance theater. Kakunin ingests behavioral events in real time, scores them against risk models, and maintains a rolling 30-day risk profile for every agent in your deployment.
Kakunin monitors: transaction (financial ops), data_access,api_call, auth_failure, scope_violation,rate_limit_hit, model_output, and custom. All events are attributed to the issuing agent's machine identity via certificate serial.
Each behavioral event contributes to a continuously-updated risk score. The engine applies weighted scoring across event types — authentication failures and scope violations carry higher weight than routine API calls. Scores are available in real time via the agent status endpoint.
High-volume event streams are passed to an anomaly detection model (Claude 3 Haiku via OpenRouter) that flags statistically unusual patterns — sudden changes in transaction frequency, novel API targets, or combinations of events that individually score low but collectively indicate compromise.
Subscribe to a live server-sent event stream of behavioral data for any agent viaGET /api/v1/agents/{id}/events/stream. Real-time governance dashboards, incident response tooling, and SIEM integrations can consume this stream directly without polling.
EU AI Act Article 12 requires logs that allow post-hoc monitoring of high-risk AI system operations. Kakunin's audit log is append-only, PostgreSQL-rule-protected, and cryptographically attributed to named agent identities — not API keys.
PostgreSQL-level audit_log_no_update and audit_log_no_deleterules block all UPDATE and DELETE operations on the audit table — even for the service role. Once written, a log entry cannot be modified or removed. Regulators can trust the log as a complete, unmodified record.
Every audit log entry carries the agent's certificate serial asactor_id. This is cryptographic attribution — the serial is issued once at registration, embedded in the X.509 identity, and verifiable independently. Log entries cannot be attributed to the wrong agent without detection.
Compliance reports and audit exports are archived to S3 with Object Lock in COMPLIANCE mode. S3 WORM provides a second layer of tamper-evidence with legal admissibility properties. Reports generated today will be verifiably unmodified when a regulator requests them in three years.
Any regulator, auditor, or counterparty can verify an agent's identity and current status via the public endpoint — no Kakunin credentials required. The verification response includes registration timestamp, issuing operator, permitted scope, and revocation history. Audit access is built in, not added later.
The ultimate AI governance enforcement action is removing an agent's operating authority. Kakunin's revocation pipeline is automated, sub-60-second, and designed to satisfy MiCA Article 72's requirement for capability withdrawal without human latency.
Behavioral risk score reaches ≥0.85. Risk engine triggers revocation pipeline synchronously. No human approval required — MiCA Art. 72 demands automated response capability.
Agent's X.509 machine identity is revoked. Serial added to public CRL. KMS key scheduled for deletion. certificate.revoked event written to WORM audit log with full context.
Registered webhook endpoints receive certificate.revoked payload with HMAC-SHA256 signature within 5 seconds. Your gateway, SIEM, and incident management tools are notified immediately.
The next request from the revoked agent fails at the verification endpoint. No TTL cache — revocation is instant. The agent is operationally blocked before any downstream service processes its request.
When a regulator asks for evidence of AI governance compliance, Kakunin generates a structured report mapping your agent deployment data against MiCA and EU AI Act requirements — on demand, signed, and S3-archived.
POST /api/v1/reports triggers an AI-assisted compliance report via the OpenRouter pipeline. The Gemini Flash model narrates risk context; the Claude Sonnet model produces the structured compliance analysis. Reports are delivered as signed PDFs via QStash async pipeline — typically under 60 seconds.
Reports include explicit section-by-section mapping to MiCA Articles 70, 72, and 75, and EU AI Act Articles 12 and 13. Each finding references the specific agent behavioral data that satisfies (or fails to satisfy) the regulatory requirement. Not a generic template — a report grounded in your deployment data.
Every report includes a risk score timeline for each agent over the reporting period, a narrative explanation of behavioral patterns, and a summary of any revocation events with their triggering scores. Compliance officers get both the data and the explanation in a single document.
Generated reports are stored in S3 with Object Lock in COMPLIANCE mode — legally tamper-evident archival. Your compliance team can retrieve any historical report years later and demonstrate to regulators that it has not been modified since generation.
AI governance infrastructure
X.509 agent identity, behavioral monitoring, automated revocation, WORM audit logs, and on-demand compliance reports. All in one API layer.