Noma Security

Overview

Noma discovers, governs, and protects AI and agents across the enterprise — from homegrown AI to SaaS agents and coding assistants. It intercepts agent actions before execution, evaluates them against policy and accumulated session context, enforces a decision, and records a tamper-evident audit trail, integrating through an SDK, a protocol gateway, an AI gateway, and an MCP gateway.

Classification

Coverage surface

MCPAPICloudSaaSData/DB

Stage

Launched

Type

Commercial

Target audience

Enterprise

Deployment

SaaSHybridSelf-hosted

Technical profile

Spec-grounded axes, verified by the TWG.

Interception architecture (R1)

Protocol GatewaySDK Instrumentation

Policy model (R3)

Hybrid

Authorization decisions (R4)

ALLOWDENYMODIFYSTEP_UPDEFER

Conformance review

Platform capabilities

• Pre-execution interception of agent actions across SDK, protocol-gateway, AI-gateway, and MCP-gateway integrations
• Session context accumulation across conversation threads and task horizons
• Policy evaluation with intent alignment at action time
• Five-outcome authorization engine: ALLOW, DENY, MODIFY, STEP_UP, DEFER
• Tamper-evident receipts with timestamp and decision context
• Cryptographic identity binding on every action receipt
• Intent-drift detection (AIDR): embedding-based scoring of each action against the session's stated-intent baseline (R7)
• OpenTelemetry export to SIEM/SOAR and observability pipelines for SOC teams (R8)
• Least-privilege enforcement via the Access Control module — constraining capabilities, autonomy, and permissions per action (R9)

Architecture

Noma satisfies all nine AARM requirements (R1–R6 core and R7–R9 extended). Core (R1–R6): Noma intercepts agent-initiated actions before execution, accumulates per-session context (prior actions and data classifications), evaluates each action against organizational policy with intent alignment, and produces one of five authorization decisions. Every decision is written to a tamper-evident audit trail with cryptographic identity binding, enabling forensic reconstruction across sessions. It integrates through an SDK, a protocol gateway, an AI gateway, and an MCP gateway. Semantic distance tracking (R7): Implemented as part of the intent-misalignment module within Noma's AIDR layer. Each session is anchored to the agent's stated intent, and every proposed action or tool call is scored for embedding-based divergence from that baseline — surfacing intent drift before misaligned-but-permitted actions execute. Telemetry export (R8): Decisions are exported via OpenTelemetry to SIEM, SOAR, and observability pipelines; a primary persona is SOC teams consuming this through Noma's SIEM/SOAR integrations. Least-privilege enforcement (R9): Operationalized through Noma's Access Control module, which constrains an agent's capabilities, autonomy, and permissions to the minimum required for each action at enforcement time, rather than provisioning blanket session-wide privilege.