Vallum

Enterprise AI agents now read files, hit APIs, send messages, and trigger actions in production systems - but security guardrails have not kept up. A single manipulated prompt can leak credentials, exfiltrate data, or trigger unauthorized actions with no audit trail. Vallum solves this with three defense layers inspired by Roman fortification engineering: SHIELD inspects every prompt in real-time using Veea Lobster Trap as a conversation-layer proxy, enhanced with Google Gemini for AI-powered intent classification. It detects prompt injection, data exfiltration attempts, and PII exposure with sub-millisecond latency, enforcing enterprise policies inline. SPEAR runs continuous automated red teaming mapped to 11 MITRE ATLAS 2026 techniques - including direct and indirect prompt injection, tool hijacking, privilege escalation, semantic jailbreak, and model evasion. A mutation engine generates payload variants using unicode substitution, base64 encoding, and roleplay wrapping to test defense robustness. CrewAI agents enable adaptive attack generation based on previous results. CHAIN provides immutable SHA-256 hash-chain audit trails with tamper-evident logging, risk scorecards, and regulator-readable compliance reports for SOC2, HIPAA, and PCI-DSS frameworks. The platform exposes a FastAPI REST service with rate limiting, a Streamlit cyberpunk dashboard connected to live data, and deploys to GCP Cloud Run with non-root containers and Secret Manager integration. Built by Team MycoGuard - an intergenerational father-daughter team of self-taught developers passionate about making AI security accessible and enterprise-ready.