Argus

Argus is an open-source, MIT-licensed security operations platform created by Thomas Unise. It collects data from cloud, identity, code, endpoint, network and Kubernetes sources, detects attacks including machine-speed ones, investigates them with AI agents, and contains them through a controlled action broker. It is self-hosted and multi-tenant.

What It Is

Argus is a self-hosted security operations platform that links assets and events from AWS, Azure, GCP, Kubernetes, GitHub, Okta, Entra, osquery and netflow into an evidence graph showing attack paths. Detection rules cover swarms, identity abuse, cloud and Kubernetes exposure, code issues, canaries, anomalies and unsanctioned AI services. Findings are enriched with CISA Known Exploited Vulnerabilities and EPSS scores, and the enrichment works offline for air-gapped installs.

How Response and Autonomy Are Governed

Every change goes through an action broker with a catalog of 27 typed actions across AWS, Azure, GCP, Kubernetes, GitHub, Okta, Entra, Cloudflare, CrowdStrike and SentinelOne. Requests pass policy checks, approvals, a dry run with blast radius, execution and stored rollback, with an audit entry at each step. A kill switch halts actions globally or per tenant.

Seven specialist AI agents triage and investigate using read-only tools and can only propose actions. Autonomous response playbooks and autonomous pentesting are off by default and need both a global switch and a tenant opt-in. Playbooks start in shadow mode, use only reversible actions, are capped, and roll back automatically unless a human keeps the change. Approve, deny and roll back buttons are available in Slack.

AI Security Features

An AI gateway proxies OpenAI and Anthropic traffic with organization-held keys, model allowlists, data-loss prevention, prompt-injection flags, budgets and a usage log. Shadow-AI discovery finds unapproved AI services from network, cloud and SSO telemetry. Canaries aimed at AI agents include planted prompt injections, agent rules files, a fake MCP server and fake LLM API keys. An attack simulator replays 8 agent-driven scenarios in an isolated sandbox tenant and reports coverage.

Setup Path

Development runs on Python 3.11 or 3.12 with SQLite and uvicorn, no Docker needed, plus a seed script for synthetic demo data. For pilots there is Docker Compose, with a Helm chart for Kubernetes and Terraform for AWS and Azure. Production mode refuses to start without PostgreSQL, Redis, encryption keys, HTTPS, an external audit sink, and SSO or TOTP. The project states it is an open-source foundation and not a certified product, and recommends starting in monitoring-only mode. The author also offers a paid installation service.