Skip to content

Tool Security Center Demo Script ​

Seven demos. Eighteen minutes. Every tool, MCP server, and integration under one security lens.

This script walks through the GhostAI Tool Security Center — the admin surface where security teams classify tools, set enforcement policies, approve risky capabilities, trace multi-tool chains, audit supply chain posture, and export evidence for HIPAA, SOC 2, and PCI-DSS programs. The narrative assumes an organization that already uses chat tools, MCP servers, OpenAI/Anthropic function calling, and internal integrations.


Pre-Demo Checklist ​

  • [ ] GhostAI instance is running and reachable
  • [ ] Admin role (or equivalent) can open Admin Dashboard
  • [ ] Tool Security tab is visible under Admin Dashboard navigation
  • [ ] Tool catalog is populated (at least a dozen entries across categories: chat, MCP, functions, integrations)
  • [ ] Sample sessions exist so chain tracing and enforcement feed show non-empty data (run a short chat with tool calls if needed)
  • [ ] Feature flags for your demo tier are on: TOOL_SECURITY_CENTER, TOOLING_POLICY_EDITOR, TOOLING_APPROVALS, TOOLING_CHAIN_RULES, TOOLING_SUPPLY_CHAIN, TOOLING_COMPLIANCE_EXPORT (names may vary by deployment; confirm with your tenant admin)
  • [ ] Browser at 1080p or higher so posture charts and data-flow diagrams render clearly

Quick sanity check ​

bash
curl -s -H "Authorization: Bearer $GHOSTAI_API_KEY" \
  "${GHOSTAI_BASE_URL:-http://localhost:8000}/api/tooling/catalog?limit=5" | python3 -m json.tool

You should see a JSON array (or paginated items) with tool_id, category, and risk_level fields.


The Scenario ​

Northwind Health operates a regulated assistant stack: clinicians use in-app chat tools, engineering uses MCP servers for repos and tickets, and product teams call OpenAI and Anthropic tools from the same tenant. The CISO wants one place to see risk, enforce PII rules, gate new tools through approval, block dangerous tool sequences, and prove controls to auditors — without slowing builders down.


Demo 1: Tool Catalog and Risk Classification ​

Duration: 2 minutes
Story: Show that every integration is inventoried with consistent risk and data labels.

Steps (Admin UI) ​

  1. Open Admin Dashboard → Tool Security → Catalog.
  2. Filter or sort by Risk level (low, medium, high, critical).
  3. Open one high or critical tool and expand Data classification (e.g., PHI, PCI, confidential, internal).
  4. Open Data flow (or Flow visualization) for that tool — show inputs, the GhostAI policy boundary, and outbound targets.

Tool categories to mention (sidebar or filters) ​

CategoryWhat it represents
chatNative chat / assistant-invoked tools
mcp_nativeFirst-party MCP tools registered on-platform
mcp_upstreamExternal MCP servers (proxied or federated)
openai_functionOpenAI-compatible function definitions
anthropic_toolAnthropic tool / tool-use definitions
a2aAgent-to-agent tool surfaces
langchainLangChain tool adapters
integrationSaaS and workflow integrations
infrastructurePlatform hooks (logging, webhooks, storage)

Expected output ​

  • Catalog lists tools with risk level, data classification, and category.
  • Data flow diagram shows direction of data (user → model → tool → external system) relative to the policy enforcement point.

Talking points ​

"This is the single inventory. Sales calls it shadow IT; we call it shadow tools until they're here. Every path — chat, MCP upstream, OpenAI functions — lands in one catalog with the same risk vocabulary."

"Critical isn't 'we don't like it.' It's tied to data sensitivity, blast radius, and whether the tool can exfiltrate or mutate production systems. The diagram makes the DLP conversation concrete for legal and clinical safety officers."


Demo 2: Security Posture Dashboard ​

Duration: 2 minutes
Story: Give the CISO a one-glance scoreboard and PII exposure by tool.

Steps (Admin UI) ​

  1. Open Tool Security → Posture (or Security posture).
  2. Point to the overall posture score (composite of policy coverage, open approvals, violations, supply chain gaps).
  3. Open PII statistics per tool — show which tools saw detections, tokenizations, or blocks in the selected window.
  4. Show risk distribution chart (count of tools per risk band).

Expected output ​

  • Posture score updates from live enforcement (typically last 24h unless you change the period).
  • Per-tool PII stats reconcile with the enforcement feed (Demo 7 narrative hook).
  • Risk distribution matches catalog filters from Demo 1.

Talking points ​

"Executives want a number they can track week over week. Security wants the drill-down. This dashboard is both — score at the top, PII and risk distribution underneath."

"When a tool spikes in PII detections, you don't wait for a ticket. You open the policy editor next and tighten scan depth or block specific entity types."


Demo 3: Per-Tool Policy Editor ​

Duration: 3 minutes
Story: Configure enforcement the way GRC expects — not buried in YAML.

Steps (Admin UI) ​

  1. From the catalog, select a representative tool (e.g., an mcp_upstream repo tool).
  2. Open Policy → Edit.
  3. Set Scan depth (e.g., shallow vs deep content inspection).
  4. Set Action on detect: tokenize, block, warn, or log_only.
  5. Set Scan direction: inbound, outbound, or both.
  6. Constrain Entity types (e.g., allow EMAIL, PHONE, block US_SSN, CREDIT_CARD).
  7. Optionally set Blocked types as a deny list override.
  8. Set Rate limit per user per hour to cap abuse.
  9. If shown, set Data residency requirement for that tool's artifacts.

Policy fields reference ​

FieldRole
scan_depthHow aggressively content is inspected
entity_typesPII / sensitive types subject to policy
blocked_typesHard-denied types regardless of default allow
action_on_detecttokenize, block, warn, log_only
scan_directioninbound, outbound, both
rate_limit_per_user_hourThrottle invocations per user
data_residencyRegion / residency constraint for outputs

Expected output ​

  • Saving policy shows confirmation and a version or audit trail entry.
  • Optional AI-assisted actions: Generate policy from natural language, Review proposed JSON, or Explain why current settings matter (if enabled for your tier).

Talking points ​

"This is runtime enforcement, not a PDF. Scan direction matters for MCP — inbound is user prompts and arguments; outbound is what the server returns to the model."

"Tokenize keeps workflows moving for low-risk teams; block is for crown-jewel tools. Warn and log-only help you shadow-mode a policy before you flip enforcement."


Demo 4: Approval Workflow ​

Duration: 3 minutes
Story: New or elevated-risk tools do not go live without security sign-off.

Steps (Admin UI) ​

  1. Open Approvals (or Pending approvals).
  2. Show a tool in pending state (submitted by a builder or auto-flagged on catalog import).
  3. Open the request — show risk summary, data classification, and suggested policy.
  4. Approve with an expiry date (e.g., 90 days for pilot).
  5. Open a second pending item and Deny with a reason (captured for audit).

Expected output ​

  • Approved tool moves to active with an expiration timestamp; renewal can re-open approval.
  • Denied tool shows blocked or restricted state with reason visible to admins and optionally to the requester.
  • Audit log entries for approve/deny with actor and timestamp.

Talking points ​

"This is how you pair agility with governance. Builders register tools; security approves with a time box. Expiry forces re-review when the integration changes or regulations shift."

"Denial with reason isn't bureaucracy — it's evidence. Auditors ask 'who said no and why?' We answer in one click."


Demo 5: Chain Rules and Tracing ​

Duration: 2 minutes
Story: Attacks and accidents are sequences — block the pattern, not just the single call.

Steps (Admin UI) ​

  1. Open Sessions or Chain trace — pick a recent session_id with multiple tool invocations.
  2. Show the tool chain for that session (ordered nodes: tool A → tool B → tool C).
  3. Open Chain rules → Create rule.
  4. Define a trigger sequence using explicit tool IDs or wildcards (e.g., read_* → external_publish_*).
  5. Set action: block, escalate, deep_scan, warn, or log.
  6. Save and show enforcement preview or last matched events.

Chain rule concepts ​

ConceptDescription
trigger_sequenceOrdered pattern; wildcards match families of tools
blockStop the chain when the pattern completes
escalateRequire human approval or secondary auth
deep_scanForce maximum scan depth from that step onward
warn / logAlert-only modes for rollout

Expected output ​

  • Session view lists tools in invocation order with timestamps and outcomes.
  • New rule appears in the rules list; evaluation shows matched / not matched for a sample session.

Talking points ​

"Single-tool policies catch naive exfiltration. Chain rules catch 'read the database, then post to Slack' — the kind of pattern red teams actually use."

"Wildcards let you encode policy once for a class of integrations instead of maintaining hundreds of literal IDs."


Demo 6: Supply Chain and Shadow Tools ​

Duration: 3 minutes
Story: MCP and third-party tools are software supply chain — treat them that way.

Steps (Admin UI) ​

  1. Open Supply chain report.
  2. Walk through transport: TLS version, certificate pinning or trust store hints, auth method (OAuth, API key, mTLS).
  3. Show schema drift: fingerprint of tool/schema last seen vs current; highlight drift detected rows.
  4. Open Behavioral flags (unexpected arguments, new endpoints, latency anomalies) if populated.
  5. Open Shadow tools — show tools observed in traffic or logs not yet in the approved catalog.
  6. Optionally expand Blast radius for one MCP server: dependent tools, downstream data classes, and user populations affected.

Expected output ​

  • Supply chain table lists MCP servers and integrations with health and drift status.
  • Shadow tools list differentiates discovered vs approved; one-click promote to catalog or block (if enabled).

Talking points ​

"MCP is a supply chain. If the upstream schema changes, your security posture changed — even if nobody opened a ticket. Fingerprints make drift visible."

"Shadow tools are the new shadow IT. We detect them from real invocations and bring them under policy before they become the next breach headline."


Demo 7: Compliance Export ​

Duration: 2 minutes
Story: Close the loop for HIPAA, SOC 2, and PCI-DSS evidence collection.

Steps (Admin UI) ​

  1. Open Compliance or Reports.
  2. Select frameworks: HIPAA, SOC 2, PCI-DSS (multi-select).
  3. Generate compliance report — show mapped controls (access, monitoring, change management, encryption in transit).
  4. Export CSV — open in a spreadsheet to show columns: tool id, policy version, approval status, violations, chain rule hits, supply chain status.

Expected output ​

  • PDF or HTML summary (if offered) plus downloadable CSV suitable for GRC tools.
  • Report references the same policy and enforcement timestamps shown earlier in the demo.

Talking points ​

"Auditors don't want screenshots of ten consoles. They want exportable evidence that ties controls to configuration and events. This report is the bridge between engineering truth and GRC workpapers."

"HIPAA is about PHI handling; PCI about card data; SOC 2 about consistent security operations. One engine underneath — framework-specific packaging on export."


Recovery Steps ​

Tool Security tab not visible ​

Symptom: Admin Dashboard loads but Tool Security is missing.

Fix: Confirm your user has admin/security role and that TOOL_SECURITY_CENTER (or tenant equivalent) is enabled for your subscription tier. Feature flags are often tier-gated — upgrade the demo tenant or enable flags in staging.

Empty catalog ​

Symptom: Catalog shows no tools.

Fix: Run onboarding sync or connect at least one MCP server / integration source. For dry demos, import the sample catalog bundle your SE kit provides, or register a single test tool via Admin → Integrations.

Posture score stale or zero ​

Symptom: Posture dashboard shows no PII stats.

Fix: Ensure traffic has flowed through tools with PII policies enabled. Lower scan_depth to shallow temporarily to generate detections on test data, or run the policy simulation API with synthetic payloads (see full lab).

Approvals queue empty ​

Symptom: Nothing to approve or deny.

Fix: Submit a new tool registration from a non-admin user, or use Simulate new tool in admin. Some environments auto-approve low-risk tools — temporarily raise default risk classification for demos.

Chain rule never matches ​

Symptom: Evaluation shows no hits.

Fix: Verify wildcard syntax matches your deployment's pattern language (often * for segment glob). Ensure session actually contains the ordered sequence; order matters.

Supply chain drift false positives ​

Symptom: Every server shows drift after an upgrade.

Fix: Re-baseline fingerprints from Supply chain → Acknowledge current schema (wording may vary). Ensure MCP servers are not load-balanced across heterogeneous backends for the same logical name.

Compliance export missing a framework ​

Symptom: PCI-DSS section empty.

Fix: Map at least one tool with PCI data classification and enable PCI reporting in tenant settings. Some frameworks require explicit opt-in for regulatory reasons.

Enforcement feed quiet during live demo ​

Symptom: Real-time events panel empty.

Fix: Open Enforcement feed with WebSocket or SSE enabled in the browser; trigger a blocked or warned call from a test user. Confirm ad blockers are not cutting the stream.


Demo Flow Cheat Sheet ​

#DemoDurationKey artifactKey moment
1Tool catalog & risk2 minCatalog + data flowSame risk taxonomy across all categories
2Posture dashboard2 minPosture score + PII by toolExecutive score + analyst drill-down
3Per-tool policy3 minPolicy editoraction_on_detect + scan_direction
4Approvals3 minApprove w/ expiry, deny w/ reasonTime-bound governance
5Chain rules & trace2 minSession chain + trigger_sequenceWildcard sequence block
6Supply chain & shadow3 minDrift fingerprints + shadow listBlast radius for MCP
7Compliance export2 minHIPAA / SOC2 / PCI CSVAuditor-ready extract
Total~17 min

Frequently Asked Questions (From Demos) ​

"How is this different from manual tool auditing spreadsheets?"

Spreadsheets go stale the day after the workshop. The Tool Security Center is wired to live catalog state, enforcement events, and supply chain fingerprints. Policies and approvals version with the platform; exports pull current truth, not last quarter's screenshot.

"Where do policies actually enforce at runtime?"

At the GhostAI policy boundary: tool invocation requests and responses are inspected according to scan_direction, scan_depth, and entity_types. Actions (tokenize, block, warn, log_only) apply before results return to the model or user. Chain rules evaluate ordered sequences across the session.

"How does MCP protection work specifically?"

MCP servers are classified as mcp_native or mcp_upstream. Transport (TLS), authentication, and schema fingerprints are tracked in the supply chain report. Policies apply to JSON-RPC payloads in both directions. Shadow MCP endpoints discovered in traffic can be blocked until approved.

"Which compliance frameworks are supported out of the box?"

HIPAA (PHI handling and access), SOC 2 (security operations and change control evidence), and PCI-DSS (cardholder data exposure and segmentation narratives). Export packs differ by framework; CSV is the common denominator for GRC ingestion.

"What AI-assisted features can admins use?"

Depending on tier: generate-policy from natural language, query the catalog and posture in chat, review a proposed policy diff, and explain-anomaly on spikes in the enforcement feed. These augment admins; they do not bypass approval for high-risk tools unless you explicitly configure that (not recommended for regulated tenants).

"Are feature flags per tier?"

Yes. Lower tiers may expose catalog and posture only; mid tiers add policy editor and approvals; highest tiers add chain rules, supply chain, compliance export, and AI assistants. Confirm your demo tenant's entitlements before promising a specific screen.

"What is blast radius analysis?"

For a given tool or MCP server, blast radius summarizes dependent tools, data classes at risk, and approximate user reach. It answers "If this integration is compromised, what else goes with it?"

"Is there a real-time enforcement feed?"

Yes — Enforcement feed streams tokenize/block/warn/log events with tool id, user, session, and rule matched (including chain rules). Use it beside the posture dashboard during incident walkthroughs.


For hands-on API steps against the same capabilities, see the Tool Security Full Lab.

SOC 2 Type II Certified | BAA Available | HIPAA-Ready