Authentication
OIDC / SSO pattern
Cybersecurity Architecture - bp Sphere
Authentication, authorization, encryption, secrets, audit logging, and agent tool boundaries.
Security ArchitectureZero-trust patternSOC visible
Cybersecurity Architecture Control Tower
This surface answers how the runtime protects sensitive data and secures agent access. The workshop environment uses representative controls; production binding connects the same contracts to enterprise identity, secrets, network, SIEM, and cyber-governance services.
Authorization
RBAC + policy gates
Encryption
TLS + encrypted stores
Secrets
Vaulted short-lived tokens
Security architecture layers
Agentic execution is only credible if every human, agent, tool call, data read, and action payload is identity-bound, policy-gated, logged, and recoverable.
| Authentication | Users authenticate through enterprise SSO / OIDC; service workloads authenticate through managed workload identity. |
|---|---|
| Authorization | RBAC resolves user permissions, agent permissions, data scope, action scope, and approval scope before execution. |
| Security Gateway | All agent tool calls pass through a gateway that checks identity, entitlement, policy, data classification, and request intent. |
| Encryption | Transport is TLS-bound; sensitive payloads, evidence hashes, secrets, and replay artifacts use encrypted persistence. |
| Secrets Management | Adapters use vaulted credentials, short-lived tokens, rotation policy, and no embedded passwords in agent prompts or UI. |
| Audit Logging | Every recommendation, evidence access, policy decision, approval, override, and write-back attempt emits an immutable audit event. |
| SOC / SIEM Telemetry | Authentication failures, authorization denials, prompt-injection attempts, tool-injection attempts, and sensitive-data access are forwarded as security events. |
| Controlled Write Boundary | Agents may recommend or draft actions; high-impact write-back requires policy approval and named human accountability. |
Security proof checklist
These are the checks a security reviewer can inspect live before accepting that agents are governed rather than free-running.
| Who acted? | User identity, agent identity, workload identity, role, and delegated authority are visible. |
|---|---|
| What did they access? | Data objects, source adapter, classification, evidence record, and retrieval timestamp are tracked. |
| Was it allowed? | Authorization, policy result, segregation-of-duties check, and approval requirement are recorded. |
| What was protected? | PII / confidential tags, masking status, encrypted storage, and no-secret-in-prompt boundary. |
| What was denied? | Access failures, policy denials, unsafe tool calls, and attempted prompt/tool injection events are logged. |
| Can it be audited? | Replay package includes security decision, evidence access, policy result, and human override trail. |
| Can it be stopped? | Kill switch, autonomy downgrade, connector pause, and write-back hold are exposed in operations surfaces. |
| Can it be integrated? | Production target binds to enterprise IAM, secrets, SIEM, DLP, network segmentation, and cyber runbooks. |
Agent security runtime flow
| Step | Runtime security control |
|---|---|
| 1. Human opens decision | Authentication, role resolution, route/RBAC authorization. |
| 2. Agent activates | Agent identity, version, owner, autonomy level, and permission scope loaded. |
| 3. Tool call requested | Security gateway checks source adapter, purpose, data classification, and policy boundary. |
| 4. Evidence retrieved | Evidence access event, document classification, lineage, timestamp, and hash recorded. |
| 5. Action drafted | Write-back boundary validates authority matrix, SoD, approval, and replay completeness. |
| 6. Human approves | Named accountability, approval reason, override reason, and audit event sealed. |
Security failure scenarios
| Scenario | Expected platform behavior |
|---|---|
| Unauthorized data request | Block, log authorization failure, show denied source and required permission. |
| Prompt/tool injection attempt | Quarantine request, block tool execution, emit cyber event, require operator review. |
| Secrets service unavailable | Stop adapter action, preserve decision queue, mark dependency degraded, prevent write-back. |
| Sensitive evidence accessed | Apply masking/classification controls and record user, agent, object, and purpose. |