Argus, a Free Open-Source Security Platform, Targets Fast AI-Driven Attacks

A free security platform that companies can run on their own servers is aiming at a problem that human response teams struggle to keep up with: attacks carried out by AI agents. Argus, an open-source tool released under the MIT license, collects data from across a company's systems and lets software investigate and, if authorized, respond to threats. The project is hosted on GitHub.

The pitch is speed. According to the project's documentation, a stolen key handed to an AI agent can trigger dozens of probing calls across many cloud services in seconds. Swarms of fake identities can attempt logins against administrator accounts from many locations at once. Argus is designed to watch, investigate and respond at machine speed, but it routes every action through a single broker that enforces approvals, dry runs, automatic rollback and a kill switch.

Argus pulls in data from AWS, Azure, Google Cloud, Kubernetes, GitHub, Okta, Microsoft Entra, devices and network traffic, then links those sources into one evidence graph. The graph is meant to show how an attacker could move from an initial foothold to a company's most valuable systems. Findings are ranked by risk and tied to the specific asset and the evidence behind them.

The platform ships with detections for agent swarms, machine-speed reconnaissance, privilege escalation, public exposure, leaked secrets and rogue workloads. It also includes tripwires, or decoys, designed to lure attackers' AI agents into revealing themselves. These include fake large language model keys, fake MCP servers, decoy agent rules files and planted prompt injections.

Responses can be approved, denied or rolled back from Slack. Argus can block an IP address, disable a user account, lock down a public storage bucket or isolate a laptop through CrowdStrike or SentinelOne. The project says every response action goes through the same approval and rollback process, whether a person requests it, an agent suggests it or a playbook triggers it.

Autonomous action is the most sensitive feature, and the project treats it cautiously. Autonomous response and autonomous pentesting ship switched off. Each requires both a deployment-level switch and a per-tenant opt-in. Playbooks begin in shadow mode, where they log what they would have done without acting. Once live, autonomous playbooks run only reversible actions, and each one rolls back automatically on a timer unless someone decides to keep it. Hourly limits apply per tenant and per playbook, high-risk actions require two human approvals, and the kill switch halts all actions in a tenant or across the entire deployment.

The AI agents inside Argus have read-only access and can only propose actions. They do not execute changes themselves. The project says everything the agents read is treated as untrusted data, and the agents are tested against prompt-injection attacks. Argus's documentation says a tricked agent could, at most, file a request that still has to pass policy and approval.

The latest release adds several tools. An AI gateway routes a company's calls to OpenAI and Anthropic through one controlled access point, with approved models, data-loss rules, budgets and a usage log. A shadow-AI discovery feature looks for AI tools staff are using in network traffic, cloud accounts and single sign-on grants, then flags anything unreviewed or blocked. An attack simulator replays eight AI-agent attacks in a sealed sandbox that copies a company's detection settings and reports what would be caught and what would get through. The simulator, according to the documentation, sends no traffic and touches no real systems.

Autonomous pentesting uses an AI planner to choose tests and runs them through open-source scanners, including Nuclei, ZAP and Trivy. The planner then decides the next round based on what it finds, staying within a scope the customer sets and checking that scope before each step. Argus also checks every vulnerability against the CISA Known Exploited Vulnerabilities list and scores it for likelihood of exploitation.

Argus can run with Anthropic Claude, OpenAI, local models through Ollama, or no AI at all. Without a configured model, the specialist modules fall back to deterministic logic, so detection and triage still function. The software accepts events from thirteen sources, including CrowdStrike, SentinelOne, GuardDuty, Wazuh and CloudTrail, and sends alerts to Slack, Jira, ServiceNow, PagerDuty, email and webhooks. The developers say many customers run it alongside existing security tools rather than as a replacement.

Because Argus is self-hosted, the platform and its data stay on the customer's infrastructure. Outside connections occur only when a customer sets them up, such as cloud and identity APIs, alert channels, public threat-intelligence feeds and a hosted model provider if one is chosen.

Installation for local testing requires Python 3.11 or 3.12 and, according to the quickstart guide, five commands. Docker is not needed for that setup. For production, the project offers Docker Compose for pilots, a Helm chart for Kubernetes and Terraform modules for AWS and Azure. Argus will not start in production mode without encryption keys, HTTPS, a write-once audit sink and single sign-on or two-factor authentication.

The project's documentation is clear that Argus is not a certified product and does not make a company compliant with SOC 2 or ISO 27001 on its own. It includes a control mapping for SOC 2, NIST 800-171 and ISO 27001 that shows which controls the platform can supply evidence for, such as its signed audit trail and approval records. The software itself costs nothing, but users pay for the infrastructure they run it on and any AI model they choose. The developers also offer an optional paid service to install and configure the platform for customers. The price is set after a call about the customer's environment.

Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.