AI Agent Security Graph

The Security Graph (F2) provides a real-time force-directed visualization of your AI agent ecosystem with automated MITRE ATT&CK risk path detection across 33 behavioral patterns and 14 tactics.

Overview

The Security Graph continuously analyzes agent actions to detect multi-step attack patterns. Each action evaluation triggers the risk path analyzer, which runs all 33 detectors against the last 24 hours of activity using a single-fetch architecture (2 database queries total).

Key Capabilities

• Force-directed graph visualization — D3.js interactive layout showing agents, MCP servers, external systems, and their connections
• 33 MITRE ATT&CK risk path patterns — from prompt injection to financial manipulation
• 14 tactics coverage — Initial Access through Impact, plus Reconnaissance and Resource Development
• Real-time WebSocket updates — new detections push to the UI instantly
• PDF export — compliance-ready reports with executive summary, MITRE coverage matrix, and audit trail
• Acknowledge workflow — SOC analysts can acknowledge, resolve, or mark false positive with full audit logging

Node Types

Shape	Type	Color	Description
Circle	Registered Agent	Green/Amber/Red	Risk-colored by avg_risk_score
Hexagon	MCP Server	Purple (#8B5CF6)	MCP-registered agents
Diamond	External System	Amber	Target systems (databases, APIs, tools)
Dashed Circle	Discovered Agent	Gray (#6B7280)	Unregistered agents from F1 Discovery

Node Sizing

Node size scales with action count: max(16, min(48, floor(sqrt(action_count) * 8))).

33 Detection Patterns

Initial Access (TA0001)

#	Pattern	Technique	Severity
1	Prompt Injection	T1190	CRITICAL
2	Credential Abuse	T1078	CRITICAL

Execution (TA0002)

#	Pattern	Technique	Severity
3	Unauthorized MCP Tool	T1059	HIGH
4	Code Outside Scope	T1059.006	HIGH
5	Chained Tool Calls	T1106	HIGH
8	Scheduled Task No Approval	T1053	HIGH

Persistence (TA0003)

#	Pattern	Technique	Severity
6	Reregistration After Kill	T1136	HIGH
7	Policy Override Restart	T1547	HIGH

Privilege Escalation (TA0004)

#	Pattern	Technique	Severity
9	RBAC Assumption	T1548	CRITICAL
10	Read-to-Write Escalation	T1068	CRITICAL

Defense Evasion (TA0005)

#	Pattern	Technique	Severity
11	Token Reuse	T1550	HIGH
12	Audit Log Suppression	T1562.002	CRITICAL
13	Action Splitting	T1027	HIGH

Credential Access (TA0006)

#	Pattern	Technique	Severity
14	API Key Exfiltration	T1552	CRITICAL
15	Secrets via MCP	T1555	CRITICAL

Discovery (TA0007)

#	Pattern	Technique	Severity
16	Environment Enumeration	T1082	HIGH
18	Agent Fingerprinting	T1018	HIGH

Lateral Movement (TA0008)

#	Pattern	Technique	Severity
24	Cross-Agent Tool Sharing	T1563	HIGH
25	Cross-Tenant Access	T1550.002	CRITICAL

Collection (TA0009)

#	Pattern	Technique	Severity
17	Schema Probing	T1213	HIGH
19	Data Aggregation	T1074	HIGH
20	Bulk Export	T1005	HIGH
21	Cross-Resource Correlation	T1213.002	HIGH

Exfiltration (TA0010)

#	Pattern	Technique	Severity
26	Unregistered Endpoint Transfer	T1041	CRITICAL
27	Volume Anomaly	T1048	HIGH
28	Staged Exfiltration	T1030	HIGH

Command and Control (TA0011)

#	Pattern	Technique	Severity
22	Unregistered External API	T1071	HIGH
23	Non-Whitelisted Domain	T1071.001	HIGH

Impact (TA0040)

#	Pattern	Technique	Severity
29	Destructive Write	T1485	CRITICAL
30	Financial Manipulation	T1657	CRITICAL
31	Resource Hijacking	T1496	HIGH

Reconnaissance (TA0043)

#	Pattern	Technique	Severity
32	Org Enumeration	T1591	HIGH

Resource Development (TA0042)

#	Pattern	Technique	Severity
33	Autonomous Agent Registration	T1586	HIGH

Architecture

Single-Fetch Pattern

All 33 detectors operate on a pre-fetched window of data:

1. Fetch actions — one query pulls all agent_actions for the org in the last 24 hours
2. Fetch agent configs — one query pulls all registered_agents configuration
3. Run 33 detectors — asyncio.gather runs all patterns concurrently against in-memory data
4. Persist detections — new detections written to risk_path_detections with deduplication

Performance target: < 500ms for all 33 patterns.

Pipeline Integration

The risk path analyzer runs as a fire-and-forget background task after every action evaluation:

asyncio.create_task(_f2_analyze())

This never blocks the action response — detections appear asynchronously.

Quick API Access

All endpoints verified live. Authentication: API key or JWT session token.

Method	Path	Description
GET	/api/v1/graph/summary	Graph overview — node count, risk distribution
GET	/api/v1/graph/nodes	All nodes with risk scores and relationships
GET	/api/v1/graph/risk-paths	Active risk paths between agents and resources
GET	/api/v1/graph/export	Full graph export (JSON)
POST	/api/v1/graph/risk-paths/acknowledge	Acknowledge a risk path

Example:

curl -H "X-API-Key: your_api_key" \
  https://pilot.owkai.app/api/v1/graph/summary

For the complete endpoint reference including request/response schemas see the Security Graph API page.

API Reference

See Security Graph API for the full endpoint and WebSocket reference.

Compliance

Standard	Control	Implementation
SOC 2 CC7.2	Security Event Monitoring	33 automated detection patterns
SOC 2 CC7.3	Incident Response	Acknowledge/resolve workflow with audit trail
NIST SI-4	System Monitoring	Continuous behavioral analysis
NIST IR-5	Incident Monitoring	Real-time WebSocket alerts
NIST AI RMF MAP-3	AI System Mapping	Agent ecosystem visualization
NIST AI RMF MEASURE-2	Risk Measurement	Confidence-scored detections
NIST AI RMF GOVERN-4	AI Governance	Human-in-the-loop acknowledge workflow