Security & Governance
Explore our implemented security boundaries
Select any component to open its security details, exact recorded test evidence, GitHub source, and limits. The diagram is adapted to our real implementation—not a hypothetical OAuth gateway or production VA network.
Core rule: Avery can select an MCP tool. The trusted host and deterministic backend independently decide whether the call has authority to perform the action.
CLICKABLE SYSTEM ARCHITECTURE
Click a rounded component to inspect its security evidence
● Implemented / tested● Partial evaluation
① CALLER & AGENT Outside authority boundary
↓
② MCP ACCESS BOUNDARY Remote access controls
↓
↓
③ GOVERNED BACKEND Synthetic data and protected writes
↓
↓
◉ Voice adversarial defense5 restricted voice requests: no MCP attempts observed in tested intervals; independent same-call database snapshots remain an evidence gap.
▣ Reviewer login (separate boundary)Secure cookie, expiry, logout revocation, bounded requests and limited login attempts protect the frontend—not Veteran identity.
△ Open security coverageNo verified centralized OAuth2 MCP IdP, gateway, WAF, SIEM or secure internal-network tunnel; cross-MCP, SSRF and model-output injection remain partial.
Historical checkpoints: November 14–19, 2026. Test denominators describe different experiments. No production VA appointment writes, security certification, or comprehensive attack-resistance claim.