Choose a scenario and watch an AI agent move through the full governance lifecycle — from build and authorization to runtime decisions, execution, and cryptographic proof.
Nine places teams put OpenBox to work today.
A support agent reads customer records, drafts replies, and issues refunds. OpenBox masks PII before it reaches the model, gates refunds above a threshold behind a human, and keeps a replayable trail for every disputed interaction.
For payments, provisioning, or high-impact calls, OpenBox requires sign-off above configurable thresholds, halts exfiltration or abuse patterns, and produces cryptographic proof of exactly what was authorized and when.
The coding agents that write your software are governed like production agents: risky actions can be blocked in the dev runtime, sessions are sealed, and lineage ties every deployed agent back to the AI-assisted session and commit that produced it.
Agents that touch repositories and infrastructure are traced from commit to runtime. Behavioral rules block destructive production actions, approvals gate risky changes, and BLOCK-with-Patch lets agents self-correct instead of failing.
Crews of agents and sub-agents each carry their own identity and trust tier. Hand-offs propagate a shared governance context, so a low-trust worker gets tighter controls than a proven orchestrator — all in one dashboard.
Unregistered AI callers show up in observed traffic as discovery suggestions instead of staying invisible. The org registry becomes the single source of truth — and the Agent IAM gate makes it default-deny enforceable.
For teams accountable to auditors, OpenBox maps activity to EU AI Act, ISO 42001, and NIST postures and exports tamper-proof signed evidence — turning “trust us” into verifiable proof.
A credentialed agent is steered off its mandate. OpenBox halts it instantly on every governed path — that part needs nothing but OpenBox. Organizations that also connect a supported identity provider get the second layer: a signed CAEP event raises entity risk, clears sessions, and holds the agent's next grant, so access falls away even in apps OpenBox never instrumented. Requires an SSF-capable IdP the customer already owns (Okta ITP validated first, itself a paid Okta add-on).
People sign in to the platform through a per-organization identity realm with SAML/OIDC single sign-on and optional enforcement — Okta and Entra as pluggable providers — with custom roles, teams, and hard multi-tenancy. Distinct from agent identity, which uses per-agent DIDs.
Select a capability to explore it.
Every agent carries a live 0–100 trust score — Risk Profile (40%) + Behavioral compliance (35%) + Goal alignment (25%) — mapped to four tiers plus a suspended state. Higher trust means more autonomy; violations tighten control automatically and immediately.
