Trust infrastructure for AI agents, models, devices and high-impact actions.
AtlasProof Control Plane helps teams govern AI-driven actions before execution — with policy gates, approvals, one-time execution tickets, evidence capsules, replay, quarantine and audit-ready exports.
AI can act now. Most governance still watches from behind.
AI agents can now move money, write records, change infrastructure and act on devices.
Most teams discover what an agent did after it happened — through logs, not gates.
Approvals, rollback and audit are bolted on per tool, per team, per vendor.
How AirTrustOS works
One policy pipeline in front of every consequential action: classify, check, approve, ticket, prove.
Classify
Every action gets an effect class, E1 (draft) to E5 (hard to reverse).
Check
Policy predicates verify identity, scope, risk and integrity — indeterminate means fail-closed.
Approve
High-impact actions pause for a human. Nothing high-impact runs unattended.
Ticket
Allowed actions get a single-use, expiring execution ticket for their rail.
Prove
Every decision produces an evidence capsule — replayable, quarantinable, exportable.
Seven active gates
Vertical trust gates share one engine contract, one control plane and one evidence layer.
FinanceGate
Payments & commerce
Execution control for AI-initiated payments, refunds, payouts and budgets.
Egifta Trust Layer
Gifting & commerce product
The trust layer securing the Egifta gifting product with the same shared flow.
ClinicalGate
Healthcare workflows
Clinician authority, reversibility windows and break-glass control for simulated clinical actions.
EduGate
Education workflows
Educator authority, student impact levels and gradebook/credential gating.
Agents, models and providers — governed before they act
Register model providers with trust evaluations, domain scopes and effect ceilings. Bind agents to providers, scope their tool permissions and gate their autonomy. Unknown providers, expired credentials and out-of-scope tools fail closed.
In the registry
- Hugging Face, AWS Bedrock & SageMaker, NVIDIA NIM (mocks)
- Self-hosted open-source models and enterprise gateways
- Connector registry across 12 system types
- Credential vault — status metadata only, never secrets
Every decision leaves evidence
- Evidence capsules for allowed, pending, rejected and fail-closed decisions
- Deterministic replay to detect policy drift
- Quarantine for risky agents, models, devices and tools
- Compliance export packs per domain and time window
Built audit-first, not audit-later
The same pipeline that gates an action also records it. Replay, quarantine and export live next to the decision — not in a separate tool.
See it decide
Full guided demo →AI agent requests payment execution
HighE4 vendor payout → pending approval → one-time ticket on the mock rail
Run in FinanceGate →Clinical assistant requests emergency access
CriticalE5 break-glass → justification required, or fail-closed with quarantine
Run in ClinicalGate →Education agent attempts grade write
HighE4 grade write → educator approval before any ticket
Run in EduGate →BYOD device requests high-impact access
HighFailed attestation → fail-closed, device workflow quarantined
Run in DeviceGate →Coding agent requests production deployment
CriticalNo rollback readiness → fail-closed; ready → release approval → ticket
Run in DevGate →Open-source model onboarding requires trust evaluation
MediumMock trust evaluation → evidence capsule → live registry entry
Onboard a provider →Sandbox-only demo
Everything on this site and in the control plane is simulated. No real payments, healthcare or education writes, device control, model-provider calls, repositories, deployments, credentials or external integrations of any kind are used.
Put a gate in front of your AI.
Open the live sandbox and run a governed action end to end — from envelope to evidence capsule — in under a minute.