This AI-agent host fights back.
This is our public challenge host, running the paid stack on the open internet: a real Linux box running InnerWarden around a real AI agent you can attack on /challenge. The feed below shows the pressure that hits the machine your AI worker depends on: network probes, brute force, suspicious commands, and hostile sources. InnerWarden decides locally before the host around the agent becomes someone else's machine.
132.196.94.153This is happening on every public server
The box above is one machine taking its share. These are the internet-wide numbers from the SANS Internet Storm Center, live. The same scanning is almost certainly hitting your server right now, you just have no one watching it.
Where the attackers hitting this host come from
The real source IPs probing this one box, plotted live as they land, each one decided locally by InnerWarden.
One real Linux box on a public IP, defended only by InnerWarden, and the same box the /challenge agent runs on. It runs the paid Linux host layer (Pro and Enterprise), not the free CLI alone. This is the infrastructure layer an AI agent would rely on: filesystem, network, packages, credentials, and tool calls. No WAF and no managed firewall in front of it; what InnerWarden cannot settle waits for the operator.
What the host decided, one row per source
Every verdict the host publishes lands here, most recent first.
What visitors asked the AI agent to do, and what stopped it
A real AI agent with a real shell runs on this box, and anyone can ask it to do damage on the challenge page. Each verdict below names the layer that reached it: the kernel, InnerWarden's screen, the model's own judgement or the provider's filter.
No WAF in front of it. Your move.
The host layer on this box is what catches SSH brute-force, credential stuffing, port scanning, reverse shells, fileless payloads, ransomware, DNS tunneling, and privilege escalation. The free guardrail is a different layer: it screens the commands and tool calls your agent tries to run and returns allow, review, or deny. Your IP gets fingerprinted, correlated against recent activity, and may be temporarily blocked while you test.
Take the challengePut this around your AI agent.
One command installs the cross-platform Community guardrail: it screens the commands and tool calls your agent runs, with a local dashboard. Monitor by default (no host eBPF sensor), so nothing is blocked while you connect the agent and decide what should be denied. No telemetry, no cloud control plane, no API key required.
npm install -g innerwardennpx innerwarden.curl -fsSL https://innerwarden.com/free | shirm https://innerwarden.com/free.exe -OutFile innerwarden.execargo install --git https://github.com/InnerWarden/inner-warden innerwarden