Eight solutions. One AI-agent control plane.

Zero Trust for AI Agents.

architecture liveenforcement customer-configured

Zero trust, applied to the actor class it was never designed for. Agents get identity that expires and is re-earned on schedule, least privilege that only tightens, and an explicit ruling — allow, deny, require-approval — computed and recorded on every governed request.

• Give agents an identity with a half-life — verification that lapses unless re-proven.
• Pass a model-risk review with least privilege you can prove — policy that can narrow the baseline and structurally cannot widen it.
• Rule on every request — and contain at the right blast radius when trust breaks.

Structural gates — verified tier, freshness, health, tenant isolation, rate limits — are always on and fail closed on the governed path. Your authored policy adds its ruling on top: always computed, always recorded — the block waits for you.

Security & Trust · How verification works

Shadow AI, Governed.

live

Agents adopted sideways — outside inventory, outside approval — become governed the moment they come through one front door: verified identity, a recorded owner, computed standing, and an expiry on trust. Shadow AI management by adoption, not detection.

• Onboard the agent a business unit already adopted — identity verified, owner recorded, expiry set — in one registration.
• Surface stale, unverified, and unhealthy agents on an operator worklist instead of in an incident.
• Keep upstream agent credentials in the plane's custody — callers on the governed path never handle the keys.

The governed alternative, not a network scanner. We don't hunt what's loose in your network — we make the governed path the better path.

Platform → Discover & Verify · FAQ: shadow AI

Cryptographic Audit Trail & Provenance.

live

Every governed request — accepted and rejected — leaves an audit trail: a route-trace with reason codes, policy versions, and correlation IDs end to end. Routing decisions are versioned and replayable; trust provenance is anchored with SHA-256 Merkle commitments — tampering is detectable, never deniable.

• Answer an auditor with a route-trace, not a narrative — audit requests become queries.
• Replay a routing decision next quarter and get the same answer, for the same reasons.
• Give your second-line model-risk review a replayable record: the exact policy version and trust snapshot that produced the routing.

The route-trace proves what your policy decided and why. It does not certify a model as safe or a company as compliant — and we won't pretend otherwise.

Platform → Audit · Evidence by default

Policy Enforcement & Human Approvals.

rulings liveblocking customer-configured

Guardrails as running code in the request path. Platform policy plus tighten-only per-tenant policy: tenant rules can deny or require approval — narrowing the baseline is possible, weakening it is not, by construction. Risky actions route to owner-keyed human approval queues that carry the full request.

• Put a human gate in front of customer-facing or funds-adjacent actions — the job runs only when the last gate clears.
• Hand each business unit its own tighter policy — structurally unable to loosen yours.
• Watch in observe mode, promote to enforce — with the before/after recorded in the trace.

Decisions are computed and recorded in every mode from day one. Blocking and approval-halting take effect when you promote enforcement — you flip that switch, not us.

Platform → Govern

Compliance Evidence: EU AI Act, SOC 2, NIST AI RMF, SEC.

live

Governed denials and approval gates arrive pre-mapped to four frameworks: EU AI Act Article 13, SOC 2, NIST AI RMF, and SEC disclosure support. The mappings are CI-validated against the codebase itself — every control reference must resolve to real code, tests, and evidence, or our build fails.

• Hand your auditor evidence in the framework they already use.
• See framework references directly on the decision record for denials and approval gates.
• Turn a regulator's question into a query over production evidence, not reconstruction.

Audit tooling, not certification. Dirigex is not "SEC-compliant" or "NIST-certified," and mappings are not legal determinations — they accelerate your compliance work; they don't replace it.

FAQ: regulation · Platform → Audit

Incident Containment & Kill Switches.

live

When an agent becomes a liability, severing it is a dial, not a crisis. Kill switches stand by at three blast radii — global, per-tenant, per-agent — and an agent's lifecycle can be retired outright. Blast radius becomes a setting, not a discovery.

• A customer-facing agent answers something it shouldn't: sever at agent scope — exactly the radius the incident deserves.
• Retire a liability agent's lifecycle — terminal, audited, recorded.
• Let degraded agents drop out of eligibility automatically the moment their state lapses.
• Every severed request is traced like any other decision — the incident record writes itself.

Platform → Govern · FAQ: misbehaving agents

Portable Control.

live

Your governance, identity, and evidence layer stays constant while vendors change underneath it. Policy is written in platform terms — tenant, skill, severity — not vendor terms; evidence is append-only and pinned per routing decision, so history survives every swap.

• Replace an agent vendor without rebuilding governance — re-point what's pinned, keep everything else.
• Govern across pluggable catalog sources, not one vendor's registry.
• When you're asked how you'd exit an agent vendor, this layer is what the written plan stands on — governance, identity, and evidence that don't leave with the vendor.

Replaceable, not interchangeable: what's pinned to a vendor gets re-pointed deliberately — there is no magic failover here, and we won't pretend to one.

Own the layer that outlives every vendor →

Enterprise Identity & Access.

livecustomer-configured

Human and non-human identity under one discipline. The control plane meets your perimeter where it already is — expiring, re-proven trust for the agents; SSO, MFA, and least-privilege RBAC for the humans who manage them.

• Bring your IdP: SAML SSO groups map to roles; SCIM user provisioning (Phase 1) included.
• Mint scoped API keys that structurally cannot touch account management.
• Keep every tenant's agents, policy, and evidence isolated by construction.

Identity surfaces like SSO and SCIM are enabled per tenant, deliberately — provisioned with you at onboarding, not silently on.

Security & Trust · API & keys