Skip to main content
platform

Every action scored. Every score explained.

A multi-layer scoring engine evaluates every AI agent action against 16 compliance frameworks in real time. Deterministic by default -- no LLM hallucination risk.

GET STARTED
0
Frameworks
0
Scoring Layers
0
Ontology Nodes
<0ms
Scoring Latency
Score decomposition
0risk score
Score Decomposition
IntrinsicAction danger level
+35
StructuralPattern analysis
+22
PolicyPolicy violations
+18
TemporalAnomaly detection
+12
Composite total0/100
Scoring in action

Every agent action receives a 1-100 composite risk score with full decomposition.

None3

Agent reads README.md

Allowed silently. No alert.

High72

Agent executes rm -rf /tmp

Blocked. Security team alerted.

Critical94

Agent claims config read, OS sees .ssh/id_rsa access

Blocked immediately. Incident created.

Four-Layer Composite Scoring

Each action receives a 1-100 risk score from four independent layers: intrinsic action taxonomy (verb + target + scope), structural pattern analysis via GNN behavioral baselines, compliance graph inference over a 1,948-node / 1,075-edge ontology with 90 rules across 7 core frameworks, and temporal anomaly detection for velocity spikes and attack trajectories.

  • Deterministic graph reasoning by default -- LLM fallback only when confidence < 0.8 for explainable justifications
  • Forward-chaining inference across 1,948 ontology nodes and 1,075 edges
  • Sub-1ms intrinsic and compliance layers; full pipeline under 50ms

Structural GNN

SubgraphGNN learns risk patterns from agent behavior graphs. 13 node types, 20 edge types, 2-layer HeteroConv architecture. Multi-task outputs: risk classification, severity regression, threat category, and binary detection. CPU inference in 5-15ms.

  • 10 threat categories: policy violation, exfiltration, privilege escalation, scope creep, excessive agency, multi-step attack, prompt injection, tool poisoning, supply chain, anomalous access
  • Multi-task heads: classification + regression + threat category + binary detection
  • CPU inference: 5-15ms per action

Divergence Scoring

When proxy intent diverges from OS truth, the divergence score amplifies the composite risk. An agent that claims to read a config file but actually accesses SSH keys receives maximum divergence penalty.

  • Cross-layer validation: proxy intent vs. EndpointSecurity truth
  • Divergence score amplifies composite risk on mismatch
  • Catches compromised, misconfigured, and malicious agents

Compliance Knowledge Base

A proprietary compliance ontology of 1,948 nodes and 1,075 edges encoding 16 regulatory frameworks. Forward-chaining graph inference evaluates 90 rules in sub-1ms, mapping every agent action to specific risk factors, regulatory articles, and mitigation recommendations.

  • 16 frameworks including GDPR, HIPAA, SOC 2, EU AI Act
  • 90 inference rules across 7 core frameworks
  • Retrieval-augmented generation from Memgraph when graph confidence < 0.8

Explainable by Design

Every score traces to specific rules, compliance articles, and behavioral signals. Full score decomposition across all four layers, built for auditors.

  • Human-readable justification for audit review
  • Full decomposition: intrinsic, structural, policy, temporal

Behavioral Engine (40+ Modules)

A production behavioral engine built on 40+ analysis modules: Markov chain transition analysis, EWMA trend detection, drift detection, group envelopes for cross-agent fingerprint merging, noise budget tracking, and Count-Min Sketch for frequency counting. Builds a 3.1KB fingerprint per agent capturing capability distributions, action frequencies, and tool usage patterns.

  • 3.1KB behavioral fingerprint per agent
  • Markov chains + EWMA for temporal anomaly detection
  • Count-Min Sketch for efficient frequency analysis
  • Group envelopes merge cross-agent behavioral fingerprints

Five-Tier Risk Bands

Every 1-100 score maps to a five-tier risk band with configurable thresholds, giving security teams a shared vocabulary from None through Critical. Thresholds are fully adjustable per tenant to match your organization's risk appetite.

  • None (1-10) · Low (11-30) · Medium (31-55) · High (56-80) · Critical (81-100)
  • Configurable thresholds -- adjust bands to your risk tolerance
  • Consistent taxonomy across alerts, dashboards, and audit exports

Secure your agents.
Ship with confidence.

One install. Every agent. Deploy in under 2 minutes. Free for your first two machines.

GET STARTED FREE