Skip to content

Tool Security Center ​

GhostAI's Tool Security Center provides comprehensive governance for every AI tool call — scanning, enforcing, auditing, and reporting on how data flows through LLM-connected tools in real time.

Overview ​

When an AI assistant calls an external tool — web search, code execution, canvas rendering, file processing, MCP servers — data leaves the chat context and traverses external services. The Tool Security Center ensures every such interaction is:

  • Risk-classified with a tool catalog that assigns risk levels based on data flow analysis
  • Policy-enforced with per-tool PII scanning, approval workflows, and configurable actions
  • Audited with per-tool-call logging of scan results, actions taken, and PII counts
  • Visible with real-time security receipts in chat and an administrative enforcement activity feed

Tool Catalog and Risk Classification ​

Every tool in the platform — built-in tools, native MCP tools, upstream MCP servers, and cloud integrations — is registered in a unified catalog with:

FieldDescription
Risk LevelDerived from data sensitivity, external exposure, and PII handling
Data ClassificationCategorized by confidentiality tier
Scan DepthControls the PII scanning thoroughness applied to tool calls
Data FlowOrdered list of hops describing where data travels
Compliance TagsFrameworks the tool maps to (HIPAA, SOC 2, PCI-DSS, GDPR, NIST, ISO 27001)

Administrators can view and filter the catalog by risk level, data classification, and approval status.

Data Flow Visualization ​

Each tool includes an interactive data flow diagram showing every hop data takes from user input to tool response. The diagram highlights:

  • GhostAI protection layer — a visual wrapper showing where the platform applies scanning
  • PII scan points — hops where data is inspected for personally identifiable information
  • External boundaries — where data leaves the platform to third-party services
  • Protection indicators — TLS status, tokenization, and encryption at each hop

Canvas and Artifact Security ​

Canvas tools generate interactive HTML artifacts that render in a secure sandboxed environment. GhostAI applies multiple layers of protection to ensure safe execution:

  • Content validation — HTML content is inspected for dangerous patterns before rendering, blocking potential sandbox escapes and data exfiltration vectors
  • PII scanning — Canvas content is scanned for personally identifiable information using an HTML-aware approach that avoids false positives in code blocks
  • Runtime isolation — Rendered artifacts execute in a sandboxed iframe with a restrictive content security policy, preventing unauthorized resource loading and cross-origin access
  • Local data flow — Canvas content never leaves the platform; all data stays within the GhostAI security boundary

Per-Tool Security Policies ​

Administrators can configure security policies per tool:

  • Scan depth — how thoroughly to scan for PII
  • Entity type rules — which PII types to scan for and how to handle each
  • Enforcement action — what to do when PII is detected (tokenize, block, warn, or log)
  • Rate limits — maximum calls per user per hour
  • Approval status — approval gating for high-risk tools
  • Blocked entity types — specific PII types that should always be blocked for a tool

Multi-Scope Policy Cascade ​

When MCP servers exist at multiple scopes (cloud, tenant, user), security policies are resolved using a cascade:

  1. Tool-level policy — if a specific policy exists for the exact tool, it applies first.
  2. Server-level policy — if no tool-level policy exists, the server-level policy is used.
  3. Scope cascade — within each level, policies resolve from most specific to broadest: user → tenant → cloud.
  4. Default observe policy — if no policy exists at any level, GhostAI applies a default observe-only policy (standard scan depth, warn action).

Floor enforcement: Regardless of cascade resolution, the effective scan depth can never be lower than the organization's configured minimum floor. A user-scoped server that would resolve to "light" scanning is automatically raised to the org minimum (e.g., "standard" or "deep").

Policy fieldResolution rule
Scan depthMAX(resolved depth, org floor)
Blocked entity typesUnion across all scopes
Enforcement actionMost restrictive wins (block > tokenize > warn > log)
Rate limitsMIN across scopes (tightest limit applies)

This ensures that user-connected tools cannot bypass organizational governance while still allowing personal customization above the floor.

Approval Workflows ​

High-risk and critical-risk tools require explicit administrative approval before users can invoke them. The approval workflow supports:

  • Administrator-initiated approvals with expiration dates
  • Conditional approvals with automatic re-review
  • Audit trail of all approval decisions

Anomaly Detection ​

The platform monitors tool usage patterns and flags anomalies:

  • Unusual PII volume — a tool suddenly handling significantly more PII than its baseline
  • Unusual call frequency — unexpected spikes in tool usage
  • New tool patterns — previously unseen tools appearing in conversations
  • Security recommendations — actionable suggestions for improving tool security posture

Supply Chain Security ​

For MCP (Model Context Protocol) tool servers, the platform provides continuous supply chain monitoring in the Supply Chain & Shadow tab.

Transport and Connection Security ​

Each MCP server is assessed for its connection posture:

  • Transport type — automatically detected as stdio (local command-line servers) or http (remote servers connected via mcp-remote bridge or direct URL)
  • TLS verification — remote servers are checked for HTTPS endpoints; non-TLS connections are flagged as a finding
  • Authentication method — detects bearer tokens (stored or injected via args), OIDC/OAuth credentials (recognized from Authorization headers in the server's args), and connected-HTTPS fallback heuristics; unauthenticated servers are flagged

Schema Fingerprinting and Drift Detection ​

Every tool's schema (name, parameters, input types) is hashed using SHA-256 on each assessment. The fingerprint is stored and compared on subsequent runs:

  • Drift detection — if a tool's schema hash changes between assessments, the tool is flagged with schema drift, indicating the upstream server may have modified its API
  • Parameter counting — tracks the number of parameters per tool from the input_schema to identify surface area changes
  • Per-server drift summary — aggregated drift count is shown alongside the server's risk score

Risk Scoring ​

Each server receives a risk score (0–100) based on:

FactorScore Impact
No TLS on remote transport+20
No authentication configured+15
Large tool surface (>20 tools)+10
Schema drift per tool+10 each
Baseline30

Health Monitoring ​

Server health is derived from connection status: connected maps to healthy, disconnected to unknown, and error to degraded. The last connection test timestamp is displayed alongside the health indicator.

Chain Rule Policies ​

Administrators can define guardrails that apply across multi-step tool call sequences:

  • Sequence-based triggers — apply policies when specific tool call patterns are detected
  • Cumulative PII thresholds — enforce limits on total PII exposure across chained tool calls
  • Configurable actions — block, escalate, or flag chains that exceed policy boundaries

Shadow Tool Detection ​

GhostAI automatically detects unregistered tool calls that bypass the managed catalog:

  • Tracks unauthorized tool usage across the platform
  • Flags unknown tools with risk assessments and recommended actions
  • Provides enrollment suggestions to bring shadow tools under governance

MCP Gateway Secrets Protection ​

The MCP Data Protection Gateway inspects tool call arguments for secrets and credentials before they reach upstream MCP servers. Administrators can configure how detected secrets are handled — from automatic redaction to hard blocking — ensuring sensitive credentials are never inadvertently forwarded to third-party services.

Enforcement Visibility and Security Receipts ​

Every tool call produces a security receipt — a structured record streamed in real time.

In-chat receipts appear inline after each tool call, showing users the enforcement action taken, PII detection summary, and scan timing. Receipts are collapsible and expandable for detailed breakdowns.

Enforcement activity feed gives administrators a searchable, filterable view of all enforcement events across the platform, with support for time-based filtering, per-tool drill-down, auto-refresh, and summary statistics.

Audit records are stored permanently for every enforcement event, capturing the tool, action, PII summary, user, session, and timestamp. These records satisfy audit control requirements across HIPAA, SOC 2, PCI-DSS, GDPR, NIST, and ISO 27001.

Compliance Evidence Export ​

Generate compliance reports mapped to specific regulatory frameworks:

  • HIPAA — PHI access controls, audit logs, encryption, transmission security
  • SOC 2 — logical access, monitoring, change management, risk mitigation
  • PCI-DSS — cardholder data protection, access control, audit trails
  • GDPR — consent records, data processing inventory, right to erasure evidence
  • NIST SP 800-171 — CUI protection, access control, audit, system integrity
  • ISO 27001 — asset management, access control, cryptography, monitoring

Reports include per-control status, weighted scoring, GhostAI mitigation evidence, and actionable remediation guidance for gaps.

Getting Started ​

  1. Navigate to Admin Dashboard → Tool Security to access the Tool Security Center
  2. Review the tool catalog and risk levels in the Overview tab
  3. Configure policies for high-risk tools in the tool detail panel
  4. Monitor enforcement activity in the Enforcement tab
  5. Export compliance evidence from the Compliance tab

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