Thomas Unise publishes Argus, an MIT-licensed open-source security operations platform

Argus is a self-hosted platform that pulls data from AWS, Azure, GCP, Kubernetes, Okta, Entra, CrowdStrike and other sources into one evidence graph. It can block an IP address, isolate a device or disable a user, with approvals handled through Slack. Autonomous response and autonomous pentesting are off by default and need a deployment setting and a per-tenant opt-in.

The AI agents in Argus investigate with read-only access and can only propose changes. Every change, whether requested by a person, suggested by an agent or triggered by a playbook, passes through one action broker with approvals, dry runs, automatic rollback and a kill switch. Playbooks start in shadow mode, and the project says its agents are tested against prompt-injection attacks.

The platform also includes an AI gateway for OpenAI and Anthropic calls, shadow-AI discovery, and an attack simulator that replays eight AI-agent attacks in a sandbox. It places AI-agent canaries such as fake LLM keys and fake MCP servers in the environment. Installation uses Docker Compose, a Helm chart or Terraform, and the project lists Claude, OpenAI, local Ollama models or no AI as model options.

Developers building agents can see how a security platform treats agent credentials, tool access and model traffic, including keeping real provider keys inside a gateway instead of in staff or agent code. The canary and tripwire approach shows one way to detect an agent that uses a leaked key. The read-only-investigate and approve-to-act split is a usable pattern for gating your own agents' write actions.