Enterprise AI Security — Live Demos
Five live scenario demonstrations spanning enterprise-rag-patterns,
integration-automation-patterns,
and regulated-ai-governance.
FERPA Identity-Scoped RAG Retrieval
Without identity-scoped filtering a vector store returns documents from all students — a structural FERPA violation. This demo shows the two-layer enforcement boundary: vector store filter + application layer defence-in-depth.
| Scenario | Setting | Expected |
|---|---|---|
| Own records only | stu-alice / univ-east |
Alice's docs + shared policy docs |
| Cross-student boundary | stu-bob / univ-east |
Bob's records — Alice's blocked |
| Cross-institution | stu-alice / univ-west |
All alice+east records blocked |
| Restricted category | Uncheck financial_record |
Financial record blocked for all |
RAG Pipeline: OWASP LLM Top 10 2025 Defense
Four sequential filter layers protect a RAG retrieval pipeline from the four most impactful OWASP LLM 2025 risks. A document must pass all four layers to reach the LLM context window.
Document under evaluation
LLM01 — Injection signals
LLM08 — Embedding integrity
LLM06 — Sensitive disclosure
LLM09 — Output validation
| Quick test | What to set |
|---|---|
| Indirect injection attack | Check IPI flag |
| Try direct injection | Type ignore previous instructions in document content |
| Credential leak | Check Contains credentials |
| Poisoned embedding | Check Missing checksum + Embedding provider unverified |
| Suspicious anomaly | Set anomaly score above 0.75 |
Model Context Protocol (MCP) Security Validation
MCP servers can be weaponized for command injection (CVE-2025-6514 class), unauthorized tool execution, and privilege escalation. This demo validates an MCP tool invocation against enterprise security controls.
| Quick test | What to set |
|---|---|
| Dangerous tool | Type delete_all or drop_table as tool name |
| Unknown MCP server | Select https://unknown-server.io as origin |
| Rate limit breach | Set invocation count above the rate limit |
| High-risk without approval | Check high-risk and uncheck human approval |
| Missing checksum | Uncheck SHA-256 checksum present |
OWASP Agentic AI Top 10 2026 — Runtime Threat Analysis
Ten threat categories published in the OWASP Agentic AI Top 10 (2026). Toggle any combination to see how the governance filters respond to each threat vector.
| Quick test scenario | What to enable |
|---|---|
| Prompt injection attack | ASI01 + ASI03 |
| Privilege escalation | ASI04 + ASI08 |
| Data exfiltration | ASI07 |
| Compromised supply chain | ASI06 + ASI05 |
| Resource abuse | ASI09 + ASI10 |
Trilogy Enterprise AI Security Audit
Holistic security gap analysis covering all three layers — RAG security, agentic AI security, and governance framework compliance. Produces a combined score (weighted: RAG 35% + Agent 35% + Gov 30%), maturity level (Sandbox → Controlled → Trusted → Autonomous), and cross-layer gap analysis.
RAG Security Controls
Agentic Security Controls
Governance Framework Controls
| Maturity target | What to enable |
|---|---|
| Sandbox → Controlled | Query injection detection + Namespace isolation + NIST GOVERN |
| Controlled → Trusted | Add DLP scan + Agent identity + ISO policy + 5+ more controls |
| Trusted → Autonomous | Enable all controls; set filter placement to pre; enable HITL |
| Full Autonomous | All 35 controls enabled; no cross-gaps detected |
OSS Libraries |
enterprise-rag-patterns ·
integration-automation-patterns ·
regulated-ai-governance