Dev.to Security πŸ” Cybersecurity πŸ‘ 0 πŸ“– 4 min read

How I stopped getting paged at 3am for raw SSH logs (and built a host agent for Linux)

If you run your own Linux boxes, you already know what /var/log/auth.log looks like five minutes after spinning up a fresh VPS. Nobody knows the IP. You haven't pointed a domain at it yet. Still, automated scanners fin

If you run your own Linux boxes, you already know what /var/log/auth.log looks like five minutes after spinning up a fresh VPS.

Nobody knows the IP. You haven't pointed a domain at it yet. Still, automated scanners find it almost immediately and start hammering port 22 with root, admin, ubuntu, postgres, and deploy.

Most security tools handle this in one of two ways:

  1. They dump thousands of raw, disconnected logs into a dashboard and leave you to piece together what happened at 3:14 a.m.
  2. They cost five figures a year and require a dedicated security engineer just to tune the rules.

I got tired of both, so I built Vigilhold (under Al's-Ops LLC) - a lightweight host agent for Linux servers and Docker hosts that watches process, network, file, and SSH signals, tells you what is actually happening in plain English, and acts only within the boundaries you set.

We killed the fake demo and put up a live lab server instead

One thing that always bugged me about security landing pages is the "interactive demo" full of canned, sanitized data where everything looks neat and tidy.

So we got rid of the demo. Instead, we put up a live public lab right on vigilhold.com.

It is an ordinary, unadvertised VPS sitting on the public internet running the actual Vigilhold agent. No login required. You can open the lab feed and watch real bots try to brute-force it around the clock (it averages around 1,600 login attempts a day from over 100 different IPs) with a 30-minute delay. You see the exact signals the agent picks up, how it correlates them, and what it blocks.

How it actually works under the hood

I didn't want a bloated agent that opens inbound ports or tries to guess what's malicious using a black-box neural network. Here is how we structured it:

  • Deterministic rules first, not AI guesswork: Decisions come from written rules and scoring, not a trained AI model. If you enable it, Claude is used strictly to write out longer plain-language investigation summariesβ€”it never decides whether to ban an IP or kill a process.
  • Privilege separation & outbound-only: The agent runs as an unprivileged user, paired with a small root helper for fixed tasks. It only makes outbound TLS connections, so you don't have to poke holes in your firewall.
  • One-command install (including Docker): You can drop it onto Ubuntu, Debian, RHEL-compatible distros, Amazon Linux 2023, Arch, or Alpine with a single curl command. If you run Docker, one docker run on the host watches both the underlying machine and every container on itβ€”no per-container sidecars needed.
  • Offline resilience: If the control plane ever goes unreachable, the agent keeps enforcing the last-synced policy locally and logs everything it does until it reconnects.

Guardrails: Observe, Assist, and Autopilot

Giving a tool permission to kill processes or lock down a production box is nerve-wracking. Every account starts in Assist mode by default:

  • Observe: Watch-only. It alerts you, and you make every call.
  • Assist (Default): Automatically hands out temporary bans to IPs actively brute-forcing SSH, web (WordPress), mail (Dovecot/Postfix/Exim), or FTP. Anything disruptive-like killing a suspicious binary running out of /tmp or isolating a hostβ€”waits for your one-click approval in the dashboard or Slack.
  • Autopilot: Acts automatically on confirmed high-confidence threats within your configured limits, while uncertain stuff still waits for a human.

Every single actionβ€”whether triggered by a rule or a teammateβ€”gets written to a hash-chained, tamper-evident audit log. And if it ever bans an office VPN range or quarantines something by mistake, bans, quarantines, and lockdowns are reversible with one click.

Pricing & trying it out

We wanted predictable pricing for small teams, agencies, and devs running their own infrastructure-no per-alert overages or "contact sales" gates just to see a number:

  • Developer ($39/mo): Covers up to 5 servers and 3 users. Real-time alerts, port-scanner trapping, SSH config monitoring, encrypted credential vault (AES-256-GCM), and a 1-year audit log.
  • Business ($149/mo): Covers up to 25 servers and 20 users. Adds Autopilot mode, behavioral anomaly baselines, CVE scanning, canary files, webhooks/integrations (Slack, PagerDuty, Datadog, Splunk), and CIS/SOC 2/PCI-DSS compliance reporting.

There is no anonymous free tier (to keep abuse down), but both plans come with a 14-day free trial with no credit card required.

If you want to see what background internet noise looks like on an unadvertised Linux box right now, check out the live lab feed on Vigilhold - no signup needed. Would love any feedback from fellow Linux admins on the architecture or what signals you wish your current setup caught faster.
Use code with caution.

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.