Argus – a free, open-source agentic security operations platform – MIT

What it is

Argus is an MIT-licensed, self-hosted security operations platform by Thomas Unise. The author also sells a scoped, fixed-price install service. The page is a product page, not documentation, so the details below are the author's claims. Code, the threat model and the go-live guide live on GitHub (docs/DEPLOYMENT.md is the named checklist).

How it works

Every change, whether requested by a person, proposed by an AI agent or fired by a playbook, goes through one action broker. Agents investigate with read-only access and can only propose. Everything they read is wrapped and labelled as untrusted, and the agents are tested against prompt-injection evaluation suites.

Events come in from thirteen sources, including CrowdStrike, SentinelOne, GuardDuty, Wazuh and CloudTrail. Alerts go out to Slack, Jira, ServiceNow, PagerDuty, email and webhooks. Device containment is available through CrowdStrike and SentinelOne. Vulnerabilities are checked against the CISA Known Exploited list and scored for exploit likelihood.

The autonomous pentesting planner runs scanners such as Nuclei, ZAP and Trivy. It re-checks the authorized scope before every step.

The attack simulator replays eight synthetic agent attacks in an isolated sandbox. It sends no traffic and cannot trigger real alerts, findings or response. Its output is a coverage report of which attacks the current rules would catch.

Preconditions / Affected

• Model options are Anthropic Claude, OpenAI, local models through Ollama, or none. With no model, the specialists fall back to deterministic logic. Agent runs are capped by steps, tokens, time and cost.
• With a local model and offline threat-intel files, no AI or intel traffic leaves the network.
• Production mode refuses to start without encryption keys, HTTPS, a write-once audit sink, and SSO or two-factor.
• Deployment options are Docker Compose (pilot), a Helm chart (Kubernetes), and Terraform for AWS and Azure.

What to do

The page describes a local trial in five commands, but does not list them. It says to open 127.0.0.1:8080, sign in with the admin login from .env, and load synthetic data with:

python scripts/seed_demo.py

The page's suggested rollout order:

1. Run in monitoring-only mode and connect only the collectors you use.
2. Tune detection rules.
3. Run playbooks in shadow mode, where they log what they would have done.
4. Enable autonomous response only if wanted. It needs both a deployment switch and a per-tenant opt-in.

The safety controls the page lists:

• Autonomous response and pentesting ship off.
• Autonomous playbooks run reversible actions only, and each rolls back on a timer unless someone keeps it.
• Hourly caps apply per tenant and per playbook, with a blast-radius limit.
• High-risk actions need two human approvals.
• A kill switch stops actions per tenant or deployment-wide.

Caveats

• The page says Argus is not a certified product.
• Core pieces (tenancy, sign-in, roles, audit trail, job queue, policy and approvals) are tested end to end on SQLite and PostgreSQL in CI. The page gives no test results for the detection rules or the AI agents.
• The control mapping for SOC 2, NIST 800-171 and ISO 27001 shows which controls Argus can supply evidence for. It does not make an organization compliant.
• The page gives no version numbers, benchmarks or independent evaluation. Prompt-injection resistance is stated as a design goal and a test practice, with no results.
• The CISA KEV prioritization and the 02:14 canary-key scenario are shown as capabilities and a walkthrough, not as measured outcomes.