Skip to content

jbdu94/ShadowAI

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 

Repository files navigation

Shadow AI Detection Framework (SAiDF)

Overview

The Shadow AI Detection Framework (SAiDF) is a behavioral investigation methodology designed to identify, analyze, and reconstruct AI-mediated execution occurring across trusted enterprise systems.

Traditional security controls focus primarily on identities, permissions, authentication events, policy enforcement, and discrete actions. As AI capabilities become embedded within browsers, SaaS platforms, copilots, workflow engines, orchestration systems, and autonomous agents, organizations increasingly face situations where every individual action is authorized, logged, and policy-compliant, yet the resulting behavior becomes difficult to explain, attribute, or govern.

SAiDF was developed to help security teams investigate these situations by reconstructing execution chains that span multiple systems, identities, workflows, and trust boundaries.

The framework is designed to complement existing security architectures, including Zero Trust, rather than replace them.


The Core Problem

Most security investigations focus on individual events:

  • Who authenticated?
  • What resource was accessed?
  • Which permission was used?
  • Which policy was triggered?

These questions remain essential.

However, AI-enabled environments introduce a different challenge:

Multiple authorized actions may combine into a larger execution path whose overall behavior is never evaluated as a single unit.

This can result in:

  • Attribution gaps
  • Behavioral ambiguity
  • Fragmented audit trails
  • Excessive trust propagation
  • Unintended downstream execution
  • Governance blind spots

SAiDF addresses this challenge by treating behavior as a sequence rather than a collection of isolated events.


SAiDF Layers

The framework consists of four investigative layers.

Identity Layer

The Identity Layer examines:

  • Human identities
  • Service identities
  • OAuth grants
  • Delegated permissions
  • Authentication events
  • Trust inheritance paths

Key Question:

Who was authorized to perform the observed actions?


Behavior Layer

The Behavior Layer examines:

  • Workflow activity
  • User interactions
  • AI-generated actions
  • Automation patterns
  • Decision propagation
  • Cross-system activity sequences

Key Question:

What behavior emerged across systems?


Data Layer

The Data Layer examines:

  • Data access patterns
  • Data movement
  • Data transformation
  • Retrieval activity
  • Content generation
  • Information propagation

Key Question:

How did information move through the environment?


Execution Layer

The Execution Layer examines:

  • Workflow execution
  • Trigger chains
  • Automation engines
  • SaaS integrations
  • API activity
  • Downstream actions

Key Question:

What execution path was ultimately created?


SAiDF Investigation Workflow

SAiDF investigations follow six stages.

Stage 1 – Detection

Initial indicators identify unusual activity.

Examples:

  • Excessive AI interactions
  • Unexpected workflow execution
  • Unusual OAuth activity
  • Cross-system automation behavior
  • Abnormal data-access patterns

Output:

Potential AI-mediated activity.


Stage 2 – Identity Analysis

Investigators identify:

  • Users
  • Service accounts
  • Applications
  • OAuth relationships
  • Delegated authority

Output:

Trust and authorization map.


Stage 3 – Data Analysis

Investigators determine:

  • What information was accessed
  • What information was generated
  • What information was transferred
  • Which systems exchanged data

Output:

Data-flow reconstruction.


Stage 4 – Behavioral Correlation

Events are correlated across:

  • SaaS platforms
  • Browsers
  • Endpoints
  • Identity providers
  • Workflow systems
  • Security monitoring platforms

Output:

Cross-system activity timeline.


Stage 5 – Execution Reconstruction

The complete execution chain is reconstructed.

Example:

User → Copilot → SharePoint → Power Automate → CRM System → External Connector

Output:

Execution-chain model.


Stage 6 – Attribution Assessment

Investigators evaluate:

  • Intent visibility
  • Trust propagation
  • Accountability gaps
  • Governance controls
  • Behavioral ownership

Output:

Attribution analysis.


Investigation Outputs

A completed SAiDF investigation produces:

  • Trust map
  • Data-flow map
  • Behavioral timeline
  • Execution-chain reconstruction
  • Attribution assessment
  • Governance findings
  • Detection opportunities

Relationship to Unmodeled Execution

SAiDF operationalizes the concept of Unmodeled Execution.

Unmodeled Execution describes behavior that emerges from chains of individually authorized actions whose combined effect is never evaluated, attributed, or governed as a single execution path.

SAiDF provides the methodology used to identify, investigate, and reconstruct these execution paths.


Intended Use

SAiDF is intended for:

  • Security Operations Centers (SOC)
  • Threat Hunting Teams
  • Incident Response Teams
  • Security Architects
  • AI Governance Programs
  • Enterprise Risk Teams
  • Zero Trust Initiatives

The framework is technology-agnostic and can be applied across cloud, SaaS, hybrid, and enterprise environments.


Research Status

SAiDF is an active research project within the Shadow AI research program.

Related research includes:

  • Shadow AI Operational Maturity Model (SAOMM)
  • Unmodeled Execution
  • Trust Inheritance Analysis
  • Attribution Gap Assessment
  • AI-Mediated Execution

License

This framework is released for research, educational, and defensive-security purposes.

Feedback, contributions, and collaboration are welcome.

About

Documents for ShadowAI maturity and toolkit

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors