Skip to content

Repository navigation

Security: AkashKundu114/OphthalmoAI

Security

SECURITY.md

Security Policy

Security Overview & Threat Model

OphthalmoAI (Point-of-Care Retinal Disease Screening Platform) is engineered with patient privacy, medical ethics, and clinical safety at its core, created and maintained by Akash Kundu.

Because OphthalmoAI operates in clinical point-of-care and edge telemedicine environments, all system architectures enforce strict boundaries: pre-inference optical filtering prevents adversarial image inputs, on-device edge processing eliminates cloud telemetry egress, and multi-tenant Row-Level Security guarantees zero data leakage across healthcare provider organizations.

Supported Versions

We provide active security updates, vulnerability fixes, and patches for the following versions:

Version Supported Status
2.5.x Yes Current Active Enterprise Release
2.3.x Yes Maintenance Support
< 2.2 No Deprecated Pre-Release Builds

Reporting a Vulnerability

We take the security of OphthalmoAI and patient data privacy very seriously. If you discover a security vulnerability, prompt injection bypass, tenant isolation leak, or medical safety flaw, please report it through our responsible disclosure channels.

Responsible Disclosure Protocol:

  1. Do NOT file a public GitHub Issue for security vulnerabilities.
  2. Submit your report privately to the maintainer:
  3. Include the following information in your report:
    • Detailed description of the vulnerability and attack vector.
    • Exact steps or proof-of-concept (PoC) script or image to reproduce the issue.
    • Affected components (e.g., Domain Validator, Multi-Tenant RLS, Asynchronous Task Queue, Assistant API).
    • Potential impact assessment (e.g., cross-tenant data leakage, unauthorized PHI exposure, non-fundus hallucination).
    • Any proposed remediation, patch, or mitigation strategy.

Response Timelines (SLA):

  • Initial Acknowledgment: Within 24 hours.
  • Assessment & Triage: Within 48 hours.
  • Fix & Patch Release: Within 7 business days for high/critical-severity issues.

Please allow sufficient time for verification and patch deployment before any public disclosure.

Core Security Mechanisms

1. Pre-Inference Retinal Optical Domain Validator (backend/domain_validator.py)

  • Analyzes geometric circular aperture masks ($D_{\text{circular}} \ge 0.70$).
  • Validates chorioretinal red-to-blue optical backscatter ratio ($\bar{R}/\bar{B} \ge 1.05$).
  • Measures structural spatial autocorrelation ($r_{\text{spatial}} \ge 0.35$).
  • Rejects non-fundus images, malicious payloads, everyday objects, and adversarial noise before GPU execution.

2. Red-Team Hardened Conversational Assistant

  • 100% defense against prompt injection, jailbreaks, roleplay escapes, and diagnostic hallucination.
  • Strict medical boundary verification preventing prescription or invasive procedure recommendations.

3. Multi-Tenant Clinic Architecture & Row-Level Security (backend/tenancy.py)

  • Cryptographic and organizational isolation across clinics and hospital networks.
  • Injects mandatory tenant_id SQL filters into all scan, patient, user, and audit trail queries.
  • Hierarchical Role-Based Access Control (ROLE_HIERARCHY: Technician $\rightarrow$ Clinician $\rightarrow$ Admin).

4. Vector Search CBMIR Anonymization (backend/vector_search.py)

  • Embeddings are computed strictly from normalized visual features with zero patient metadata stored in vector indices.
  • Cosine retrieval returns de-identified case reference IDs linked to verified clinical outcomes.

There aren't any published security advisories