SaaSGraph

OAuth & Third-Party SaaS Exposure · Hybrid Behavioral ML

A defensive SaaS-security workbench that combines explainable OAuth policy, behavioral anomaly detection, and application-to-user-to-resource blast-radius analysis.

OAuth scopes · consent · token persistence · Isolation Forest · API anomalies · access graphs · blast radius

---

Overview

SaaSGraph answers two related questions:

If a third-party application or OAuth token is abused, what can it actually reach?

Does this integration look unusual compared with a normal SaaS/OAuth reference population?

The deterministic security engine and ML layer remain separate and explainable.

flowchart LR
    A[OAuth / SaaS inventory] --> P[Scopes + consent + publisher + token state]
    A --> B[API behavior + dormancy + user reach]
    A --> G[App → user → resource graph]
    P --> R[Transparent rule score]
    B --> M[Isolation Forest]
    R --> H[Hybrid review priority]
    M --> H
    G --> H

Loading

Synthetic baseline

The public fixture contains 8 deterministic synthetic OAuth applications.

Measure	Baseline
Applications	8
Expected rule outcomes matched	8 / 8
Critical	2
High risk	2
Review	2
Normal	2
Mean deterministic risk	50.5 / 100
Users behind high-risk / critical grants	771
Highest deterministic risk	AnalyticsSync — 100 / 100

These values are synthetic implementation evidence, not measurements from a production SaaS tenant.

Hybrid behavioral ML

SaaSGraph now uses Isolation Forest over 11 OAuth/SaaS posture and behavior features:

scope risk
scope count
sensitive scope count
user reach
persistent token
application dormancy
observed / baseline API ratio
resource reach
publisher verification
admin consent
external-tenant boundary

The model is fit against 800 deterministic synthetic normal-reference profiles generated with a fixed seed.

For each application it returns:

• anomaly_percentile — unusualness relative to the synthetic normal reference;
• ml_outlier — Isolation Forest outlier state;
• top_deviations — strongest feature deviations from the normal reference;
• hybrid_priority — deterministic OAuth risk plus a bounded ML adjustment.

The existing rule score remains visible and authoritative for explicit policy findings. The anomaly percentile is not a probability of compromise.

Detailed methodology: docs/ml-anomaly-model.md.

Why rules + ML?

OAuth posture contains facts that should remain deterministic:

• sensitive scopes;
• administrative consent;
• publisher verification;
• persistent refresh-token access;
• cross-tenant trust;
• dormancy.

ML adds value when the combination of access and behavior is unusual, particularly when current API activity differs sharply from ordinary application behavior.

OAuth policy evidence
        +
Behavioral anomaly score
        +
Graph blast radius
        ↓
Explainable review priority

Deterministic OAuth risk model

The transparent rule engine still scores:

scope sensitivity
publisher trust
admin consent
token persistence
user reach
dormancy
API-volume deviation
resource reach
cross-tenant exposure
high-risk interactions

It returns:

NORMAL / REVIEW / HIGH_RISK / CRITICAL

Representative synthetic applications include AnalyticsSync, MigrationAssistant, SupportTools, LegacyExport, BuildBot, PayrollConnector, NotesLite, and CalendarHelper.

OAuth blast radius

SaaSGraph models access as a relationship rather than a flat permission list:

Third-party application
      ├── OAuth scopes
      ├── authorized users
      ├── persistent token
      └── SaaS resources
                ↓
          blast radius

For example, a broad file-reading integration may connect one publisher to hundreds of users and several downstream data systems. The graph layer makes that reach explicit while the ML layer evaluates whether the integration's posture and behavior are unusual.

Report output

The CLI now includes both deterministic and ML evidence:

{
  "summary": {
    "applications": 8,
    "expected_outcomes_matched": 8
  },
  "ml": {
    "model": "IsolationForest",
    "normal_reference_samples": 800,
    "features": ["..."]
  },
  "ml_findings": [
    {
      "app_id": "app-...",
      "rule_score": 0,
      "anomaly_percentile": 0.0,
      "hybrid_priority": 0,
      "top_deviations": ["..."]
    }
  ]
}

The ML values above are schematic field examples; executable values are generated from the checked-in synthetic fixtures.

API & dashboard

pip install -e '.[api]'
uvicorn saasgraph.api:app --reload

Endpoints:

GET /healthz
GET /report
GET /apps
GET /apps/{app_id}
GET /docs

Quick start

git clone https://github.com/VinayK88/SaaSGraph.git
cd SaaSGraph
python -m venv .venv
source .venv/bin/activate
pip install -e '.[api]'
saasgraph
python -m unittest discover -s tests -v
uvicorn saasgraph.api:app --reload

Docker:

docker build -t saasgraph .
docker run --rm -p 8000:8000 saasgraph

Portfolio distinction

BrowserGuard → browser/extension behavioral anomaly detection
AgentAtlas   → AI-agent identity/posture anomaly + peer deviation
DeepTrace    → NLP narrative clustering with TF-IDF + DBSCAN
SaaSGraph    → OAuth/SaaS behavioral anomaly + consent/access graph

Evaluation boundary

Everything in this repository is synthetic and defensive. SaaSGraph does not authenticate to real tenants, retrieve refresh tokens, enumerate private user data, revoke applications, or change consent.

The deterministic score and ML anomaly percentile are prioritization aids, not breach probabilities. Production use would require authorized tenant telemetry, role-aware peer groups, model/feature drift monitoring, analyst feedback, calibrated thresholds, and auditable response workflows.