Security &
Governance.
In-process runtime safety, fail-closed authorization architecture, tenant isolation boundaries, and privacy-first BYOK credential protection in Shree AI OS v1.0.6 Developer Preview.
Release Lifecycle
Security patches, static analysis audits, and vulnerability fixes are actively provided for the active Developer Preview:
| Version | Supported | Status |
|---|---|---|
| 1.0.6-developer-preview | ✅ Yes | Active (Current Release) |
| 1.0.5-developer-preview | ❌ No | Deprecated |
| Earlier releases | ❌ No | Discontinued |
Fail-Closed Authorization
All tool calls, subprocess invocations, and graph steps pass through graphPermissionManager. If any evaluation encounters an exception, NPE, or unmapped capability, the gate strictly returns PermissionDecision.DENY. The platform never fails open under abnormal conditions.
Strict Multi-Tenant Isolation
Every runtime request carries an authenticated TenantContext. Memory stores, session caches, reflection records, and pgvector tables enforce strict tenant_id partitioning, preventing cross-tenant data access.
BYOK Credential Protection
API keys managed through SettingsSDK reside strictly in volatile in-process JVM memory. Credentials are automatically masked before logging, never serialized to disk, and never broadcast over the event bus.
Deterministic Pre-LLM Guardrails
LLMs are treated as probabilistic generators, never execution authorities. Goal decomposition, DAG planning, capability validation, and patch safety audits occur deterministically in Java before any model invocation.
Reporting a Vulnerability
If you discover a potential vulnerability or boundary bypass in Shree AI OS, please do NOT disclose it publicly. Submit all reports via GitHub Private Vulnerability Reporting under the Security tab of the repository.
< 48 Hours
< 5 Days
< 72 Hours
Coordinated
Zero-Trust Architecture.
Production-Hardened JVM.
Review the complete platform architecture, 10 SDK facades, or run the test verification suites.