Detect

AI-BOM + Session Evidence
Detection & Investigation Engine

Use asset inventory, behavioral baselines, and D01-D06 signals to locate anomalies, then follow the session trace across identity, data, tools, and egress impact.

AI-BOM Asset Graph

Auto-discover models, toolchains, MCP, and Agent Runtime dependencies with version diff and compliance inventory.

4

Models

claude-4-sonnet

gpt-4o

deepseek-r1

qwen-2.5

23

Tools

ReadFile

EditFile

BashExec

WebFetch

+19

3

MCP Servers

github-mcp

jira-mcp

slack-mcp

5

Agent Runtimes

cursor-agent

cli-agent

custom-bot

+2

Initial Detection Rules D01-D06

Behavioral baseline drift detection that catches anomalies before incidents.

D01HIGH

New Tool First Seen

Unknown tool call beyond baseline

Depends on: ToolCall

D02MEDIUM

New Model First Seen

Unknown model load beyond baseline

Depends on: ModelLoad

D03HIGH

New MCP Endpoint

Previously unseen MCP service connection

Depends on: MCPActivity

D04CRITICAL

Sensitive Path Anomaly

Sensitive file access spike or abnormal combination

Depends on: FileAccess + ToolCall

D05CRITICAL

Permission Oscillation

High-frequency deny→allow switching, suspected bypass

Depends on: PermissionDecision

D06HIGH

Call Intensity Spike

Abnormal tool call frequency surge in a single session

Depends on: ToolCall/ToolResult

From Detection Signal to Actionable Investigation

The investigation path separates factual evidence, affected scope, and response options that are actually available.

Signal

D04 Sensitive Path Anomaly

Credential-file access combined with a new egress target

Evidence

Session Trace

Identity -> ReadFile -> BashExec -> unknown.io

Impact

Scope Confirmed

1 identity, 2 files, 1 endpoint

Response

Execute by Availability

Audit is available; other controls ship after gates pass

Baseline + Drift

Build call distribution, latency, rejection-rate, and outbound-pattern baselines per tenant/project/agent, then roll them forward after a configurable controlled learning period.

Configurable

Learning Period

Rolling

Update Frequency

tenant/agent/env

Granularity