Docs: Threat model
Short, practical risks and mitigations for onboarding ops workflows.
How we're different
- Evidence-first workflows (not chat transcripts).
- Policy-gated actions with human approvals.
- Audit trail and exports by default.
Trust rule: AI drafts; humans decide for sensitive outcomes.
Primary data flows
High-level view. Exact deployment varies by environment.
Intake
Storage (docs + hashes)
Extraction (cited)
Policy engine
Approvals
Binder export
- Documents are hashed on upload; binders include provenance.
- LLM output is constrained to structured extraction and drafting under policy gates.
- Every sensitive action requires explicit approval.
Top risks
Prompt injection / malicious documents
Documents can contain instructions that try to override policy or induce invented facts.
Unauthorized access
Onboarding packets include PII; role design and access controls must be explicit.
PII leaks (logs + exports)
Debug logs and ad-hoc exports are common leak paths in early products.
Workflow tampering
Changing state or approvals without trace undermines auditability and trust.
Mitigations
Strict policy gates
- External comms require approval
- Final outcomes require approval
- Versioned policies + change control (roadmap)
Evidence-first schema
- “No citation, no field” enforcement
- Low-confidence extractions routed to review
- Binders export citations + provenance
Least privilege + audit
- Role-based access controls
- Append-only audit timeline
- Tamper-evident exports (roadmap)
Data minimization
- Store only what is needed to run workflows and prove evidence
- PII redaction in logs (roadmap)
- Retention + deletion policies
Related
- Trust center for posture pages.
- Security for controls overview.
- Audit trail docs for the event model.
Not legal advice. Requirements vary by jurisdiction and institution.