Argus open-sources an agentic security operations platform built for machine-speed attacks

What launched

A developer named Thomas Unise has released Argus, a security operations platform that puts AI agents in charge of watching cloud accounts, logins, code and devices around the clock. The project is MIT-licensed, free of seat or usage fees, designed to be self-hosted, and reached the Hacker News front page on 24 October 2026. Its pitch responds to a shifting threat model: attackers now automate with AI, and a human-paced security team cannot out-react a machine.

According to the project's site, an agent working from one stolen key can probe twenty cloud services with sixty calls before an analyst has finished reading the first alert. Argus is built to detect, investigate and, when permitted, respond on that same timescale.

What the platform claims to do

Argus ingests data from AWS, Azure, GCP, Kubernetes, GitHub, Okta, Microsoft Entra, endpoints and network traffic, then links it into a single evidence graph meant to expose attack paths before an adversary walks them. Its detections target AI-era behavior: identity swarms, high-speed reconnaissance, privilege escalation, accidental public exposure, leaked secrets and rogue workloads. It also plants traps designed for automated attackers — fake LLM credentials, decoy MCP servers and planted prompt injections — so that an agent which takes the bait identifies itself immediately.

On the response side, the project lists 27 prebuilt actions spanning AWS, Azure, GCP, Kubernetes, Okta, Entra and CrowdStrike, including blocking an IP, isolating a laptop through CrowdStrike or SentinelOne, disabling a user and locking down an exposed storage bucket, all controllable from Slack.

Two features address AI risk inside the company rather than outside it. An AI gateway fronts all OpenAI and Anthropic traffic with centrally held keys, model allowlists, data-loss rules, budgets and prompt-injection flagging, so employees use Argus-issued keys while real provider credentials stay locked away. A shadow-AI discovery feature scans network traffic, cloud accounts and SSO grants to inventory which AI tools staff actually use and raises alerts for anything unreviewed.

For validation, Argus ships a simulator that replays eight agent-driven attack scenarios, from swarms spraying passwords to a rogue agent inside Kubernetes, in a sealed sandbox, with slower and quieter evasion variants to expose detection gaps. An autonomous pentesting mode lets an AI planner chain scanners such as Nuclei, ZAP and Trivy and iterate on what it finds, within a scope that is re-checked before every step. Vulnerabilities are prioritized against the CISA Known Exploited list, and every action lands in an audit trail held in write-once storage, with built-in control mappings for SOC 2, ISO 27001 and NIST 800-171.

How the autonomy is contained

The design choice most likely to interest teams wary of agents in production: the AI never touches systems directly. Agents investigate with read-only access and are limited to proposing changes. Every action, whether requested by a human, suggested by an agent or fired by a playbook, passes through a single broker that applies approvals, dry runs and rollback. Autonomous response and autonomous pentesting both ship switched off and require a deployment-level switch plus a per-tenant opt-in. Playbooks begin in shadow mode, logging what they would have done so operators can judge them before granting authority. Autonomous actions are restricted to reversible ones that undo themselves on a timer unless a person keeps them, bounded by hourly caps per tenant and per playbook, a limit on blast radius, and two human approvals for high-risk actions. A kill switch halts every action in a tenant or across the deployment instantly, and the project says all input the agents read is treated as untrusted and tested against prompt injection.

Running it yourself

Deployment options include Docker, Helm and Terraform modules for AWS and Azure. The platform works with Claude, OpenAI or a local Ollama model, and can also run with no AI at all. The repository is on GitHub, and the quickstart is a clone, a Python virtual environment and a local uvicorn server.

Why it matters

Every capability above comes from the project's own page, surfaced through Hacker News; there is no independent evaluation in the source material, and a tool that both observes and acts on production infrastructure deserves its own security review before deployment. That caveat aside, Argus is a notable data point in two converging trends. The first is defensive automation: as offensive AI drives down the cost of machine-speed attacks, response latency becomes the deciding factor, and a zero-cost open-source option matters to teams priced out of commercial security operations platforms. The second is agent governance: read-only investigation, propose-only agents, a central action broker, shadow mode, reversible-by-default actions and a global kill switch add up to a template for deploying autonomous agents safely in any domain, not just security. Whether the implementation lives up to the design will depend on adoption and outside scrutiny.