Your investigation evidence deserves stronger protection than ordinary application data.
SURAGX is architected with layered controls around identity, tenant isolation, hostile evidence processing, evidence integrity, AI tooling, auditability and the software supply chain.
Architecture baseline — current release verification required before claiming implementationEvidence is untrusted input. Access is scoped. AI cannot expand its own authority.
LEAST PRIVILEGETenant
Evidence
Workers
AI
Proof
Security from login to proof.
These controls are taken from the approved LLD architecture baseline. The public wording uses “designed” rather than “certified” because actual release verification must be performed against the authoritative repository and deployed build.
Keep credentials and tokens out of browser JavaScript.
- Keycloak as identity provider/broker.
- Customer SAML/OIDC federation and MFA via IdP policy.
- Authorization Code + PKCE terminated in the FastAPI BFF.
- Refresh tokens remain server-side; secure HttpOnly SameSite cookies + CSRF protection.
- Step-up authentication designed for original-evidence export/delete and role changes.
Tenant boundaries at more than one layer.
- RBAC plus organization/customer/case scope.
- Every resource lookup authorizes membership before data access.
- PostgreSQL Row-Level Security as a second enforcement layer.
- No cross-tenant deduplication, even when evidence hashes match.
- Cross-tenant BOLA/IDOR tests are part of the security baseline.
Original evidence stays original.
- SHA-256 computed server-side at ingest.
- Original evidence object is content-addressed and never overwritten by product workflows.
- Derived artifacts are separately hashed and linked to parent evidence/run.
- Every access/export/delete/processing materialization is security-audited.
- Proof Ledger uses hash-chained canonical entries for tamper evidence.
Parsers live inside a constrained worker boundary.
- Evidence is never parsed inside the API/BFF process.
- Workers run non-root with read-only evidence mount and separate writable scratch/output.
- CPU, memory, PID and time limits.
- Default-deny outbound network; no Docker socket, privileged mode or host networking.
- Archive traversal, symlink and decompression protections.
AI can investigate, but cannot silently take control.
- Allowlisted structured investigation tools only.
- No shell or command execution.
- No arbitrary SQL/OpenSearch DSL or unrestricted file-path reads.
- No direct evidence mutation, case deletion, containment or endpoint action.
- Tool calls are schema-validated, tenant-bound and server constrained.
Observe the system without leaking the investigation.
- Operational telemetry excludes raw evidence and original file bytes.
- Full prompts, access tokens, API keys and report content are excluded from normal telemetry.
- Password/API-key fields are redact-by-default.
- Security audit and Proof Ledger are separate from application logs.
- Support bundles must exclude original evidence and secrets unless securely exported by an authorized administrator.
A resource ID alone never grants access.
LAYERS
Architecture layers include application authorization, PostgreSQL RLS, object-storage scope, search ownership, processing mappings, AI tool context and privacy-safe logs.
Security controls extend beyond runtime.
The baseline requires security and supply-chain gates around every release, especially for changes to authentication, RLS, evidence upload, parser sandboxing, AI tools and the proof ledger.
The application Proof Ledger is designed to be tamper-evident. It is not represented as certified WORM storage. Legal/regulatory WORM requirements require a storage backend with independently validated object-lock/retention capability.