Quantum access should not be governed by a shared provider token, a lab workstation account, or a notebook that happens to work. Quantum systems combine expensive resources, sensitive intellectual property, export-sensitive workflows, experimental hardware, and evidence that may later support claims. Access control must be explicit and continuously evaluated.
NIST SP 800-207 describes zero trust as a model that moves away from implicit trust based on network location and toward resource-centric protection. CISA's Zero Trust Maturity Model adds a practical maturity framing across identity, devices, networks, applications, data, and cross-cutting capabilities. A quantum platform should apply these ideas to cloud targets, lab systems, control infrastructure, evidence stores, and automation agents. [R153] [R154]
View diagram source
flowchart LR
Principal[Principal] --> Decision[Policy decision point]
Device[Device posture] --> Decision
Workload[Workload manifest] --> Decision
Target[Target sensitivity] --> Decision
Data[Data classification] --> Decision
Decision --> Enforcement[Policy enforcement point]
Enforcement --> Resource[Quantum resource]Quantum resources are not homogeneous
Access to a simulator is not the same as access to a live QPU. Access to a cloud QPU is not the same as access to pulse controls. Access to evidence may be more sensitive than access to raw execution.
View diagram source
flowchart TB
Access[Quantum access] --> Simulator[Simulator]
Access --> Runtime[Cloud runtime]
Access --> Lab[Lab control plane]
Access --> Pulse[Pulse/control primitives]
Access --> Evidence[Evidence and results]
Access --> Admin[Provider and billing admin]Use resource classes.
| Resource class | Risk |
|---|---|
| public simulator | low cost, low device risk |
| managed QPU runtime | cost, queue, data, evidence risk |
| reserved QPU session | denial-of-capacity, commercial risk |
| pulse/control access | device safety and validity risk |
| calibration systems | fleet-wide reliability risk |
| evidence store | IP, customer, scientific, and audit risk |
Access is a runtime decision
Zero trust for quantum is not “login once and run anything.” Admission must evaluate context.
View diagram source
sequenceDiagram
participant User
participant Broker
participant PDP as Policy decision point
participant PEP as Enforcement point
participant QPU
User->>Broker: submit workload
Broker->>PDP: principal + workload + target + data
PDP-->>Broker: decision and obligations
Broker->>PEP: scoped token + obligations
PEP->>QPU: execute allowed operation
PEP-->>Broker: audit eventObligations may include:
- maximum spend;
- allowed targets;
- required evidence level;
- prohibition on pulse access;
- retention period;
- approval before external sharing;
- mandatory simulation before hardware execution.
Runtime isolation
Isolation should separate tenants, projects, evidence, credentials, and provider contexts.
View diagram source
flowchart TB
TenantA[Tenant A] --> BrokerA[Broker namespace A]
TenantB[Tenant B] --> BrokerB[Broker namespace B]
BrokerA --> VaultA[Secrets A]
BrokerB --> VaultB[Secrets B]
BrokerA --> EvidenceA[Evidence A]
BrokerB --> EvidenceB[Evidence B]
BrokerA --> Provider[Provider account boundary]
BrokerB --> ProviderProvider-side isolation may be weaker than enterprise isolation. The platform broker must compensate with project-level accounts, scoped credentials, tags, logs, and evidence partitioning.
Capability-based access
Do not grant access by broad role alone. Grant capabilities.
View diagram source
flowchart LR
Role[Role] --> Capabilities[Allowed capabilities]
Capabilities --> Compile[compile]
Capabilities --> Simulate[simulate]
Capabilities --> Submit[submit hardware job]
Capabilities --> Reserve[reserve capacity]
Capabilities --> Pulse[modify pulse]
Capabilities --> Publish[publish claim]Example capability policy:
capabilities:
researcher:
- simulate
- submit_research_job
- read_own_evidence
platform_sre:
- drain_target
- view_health
- open_provider_case
calibration_engineer:
- run_calibration
- approve_target_profile
external_collaborator:
- simulate
- read_shared_evidence
forbidden_combinations:
- [external_collaborator, pulse_modify]
- [billing_admin, publish_claim]Token design
Long-lived provider tokens are operational debt.
View diagram source
flowchart LR
Identity[Identity provider] --> Broker[Quantum broker]
Broker --> Token[Short-lived scoped token]
Token --> Provider[Provider runtime]
Broker --> Audit[Audit log]
Token --> Expire[Automatic expiration]Tokens should be short-lived, project-scoped, target-scoped, and bound to a workload manifest. When the provider does not support all of those constraints, enforce them at the broker.
Lab access and physical presence
Lab systems require stronger controls than remote notebook access.
View diagram source
flowchart TB
LabAccess[Lab access request] --> Identity[Identity check]
LabAccess --> Training[Training status]
LabAccess --> Safety[Safety authorization]
LabAccess --> Change[Approved change ticket]
Identity --> Permit{Permit?}
Training --> Permit
Safety --> Permit
Change --> Permit
Permit -- yes --> Session[Time-bound lab session]
Permit -- no --> Deny[Deny and record]The same person may be allowed to run cloud jobs but not adjust control electronics or cryogenic procedures.
Data-plane isolation
Execution metadata can reveal sensitive information: problem class, molecular target, optimization model, customer name, budget, and research direction.
View diagram source
erDiagram
PROJECT ||--o{ WORKLOAD : owns
PROJECT ||--o{ SECRET : uses
WORKLOAD ||--o{ JOB : submits
JOB ||--o{ RESULT : produces
RESULT ||--o{ EVIDENCE_PACKAGE : supports
PRINCIPAL ||--o{ AUDIT_EVENT : triggersProtect metadata with the same seriousness as source code.
Break-glass operations
Emergency operations must be possible, but visible.
View diagram source
stateDiagram-v2
[*] --> Normal
Normal --> BreakGlassRequested
BreakGlassRequested --> Approved
Approved --> ElevatedSession
ElevatedSession --> AutoExpire
AutoExpire --> Review
Review --> Normal
BreakGlassRequested --> DeniedBreak-glass sessions should require:
- reason;
- approver;
- time limit;
- session recording where practical;
- post-action review;
- automatic revocation.
Zero trust failure modes
View diagram source
flowchart TB
Failure[Failure mode] --> SharedToken[Shared provider token]
Failure --> NotebookSecret[Secret in notebook]
Failure --> OverbroadRole[Overbroad role]
Failure --> UnlabeledData[Unlabeled data]
Failure --> MissingAudit[Missing audit trail]
Failure --> AutomationRunaway[Automation outside policy]The remedy is not more meetings. The remedy is policy evaluation in the submission path.
Minimal quantum zero-trust checklist
| Control | Required behavior |
|---|---|
| identity | every run has a principal and project |
| device posture | high-risk operations require trusted device or lab station |
| workload manifest | target, data class, cost, and evidence level are declared |
| target policy | sensitive targets require approval |
| data policy | source/results/metadata are classified |
| audit | every decision and enforcement event is recorded |
| revocation | access can be disabled centrally |
Quantum access is not a login. It is a continuously evaluated contract.