Project brief

An autonomous SOC analyst on Amazon Bedrock — with a human on every trigger.

NyxAI is a multi-agent platform that automates security operations — alert triage, threat investigation, and incident response — while keeping human analysts in control of every real action. Deployed and running on AWS today.

The problem

SOC teams drown in alerts; time is finite.

Most security work is repetitive and labor-intensive — triaging thousands of alerts, looking up indicators across many tools, and drafting reports before any action is taken. Response is slow, 24/7 coverage is hard to staff, and real threats slip through the noise. This is exactly the load AI can lift.

What NyxAI is

Two agents, split by permission

An orchestrator routes each request to one of two permission-separated agents. NyxRead analyzes — logs, threat intelligence, GuardDuty / Security Hub findings, IAM privilege-escalation audits, XDR data — with no write access. NyxRespond executes response, but only after approval.

Humans in control

The AI proposes; a human decides

Every destructive action — blocking an IP, adding an IOC, closing an incident — is held at a human-approval gate over Telegram (NyxGuard). Authority is enforced outside the AI prompt: nothing touches a live system without an explicit Approve.

Automation

A SOAR pipeline that runs 24/7

NyxFlow continuously pulls detections from XDR, enriches indicators, and proposes an evidence-backed response for approval — then blocks across platforms on approval, or closes as a false positive. Analysts also work interactively through a web portal with SSO + MFA.

Integrations

The tools you already run

Cortex XDR, Trend Vision One, AWS WAF, Cloudflare, GuardDuty, and Security Hub — reading findings and, on approval, pushing blocks and indicators of compromise.

Built on AWS

Every workload runs on AWS

The core is Amazon Bedrock AgentCore + Claude models; each agent has its own least-privilege IAM role, and infrastructure runs in a dedicated isolated VPC. The largest and fastest-growing cost is Amazon Bedrock inference, scaling with alert volume, integrations, and analyst seats.

Amazon Bedrock · AgentCore Claude models LambdaEC2CloudFrontAPI GatewayDynamoDBCognitoS3Route 53SESSecrets Manager
Security of the AI

Assessed against frameworks

Evaluated against OWASP LLM Top 10, OWASP Agentic ASI, MITRE ATLAS, and STRIDE — with structural read/write separation, human-in-the-loop control, JWT + MFA, input guardrails, and an isolated VPC.

Roadmap

Where it's going

A model router (cheap open-weight models for routine tasks, higher-tier Claude for complex reasoning) to optimize cost; optional self-hosted open models for fully in-account inference; and broader SOAR integrations.

Architecture

How the project fits together

Entry Gateway AI core · Bedrock Agents Integrations AnalystsWeb portal · SSO + MFA Automated SOARscheduled · 24/7 Web gatewayCloudFront · Cognito · API GW · Lambda SOAR runtimen8n on EC2 · proxy Lambda NyxOrchestratorAmazon Bedrock AgentCore + Claude · guardrail NyxReadanalysis · reporting · no write NyxRespondblock · IOC · close incident ⚑ NyxGuard · human approves Cortex XDR Trend Vision One AWS WAF Cloudflare GuardDuty Security Hub
Analysts and the scheduled SOAR pipeline both reach the orchestrator on Amazon Bedrock. The read path runs freely; the write path passes NyxGuard — a human approval — before touching any live security tool. Every component runs on AWS.
At a glance
Status · deployed & running Website · nyxai.click Core · Amazon Bedrock + Claude Region · us-east-1 Control · human-in-the-loop on every action