AI Agents Are Attacking Web Shops. What Shows Up in Your Logs?
In October 2026 researchers at NorForge described an attack tool called Cairn: autonomous AI agents that went after more than 1,200 web shops, succeeded in roughly one in ten, and spent about $50 per successful compromise. The goal was a web shell, a small script on the server that gives the attacker lasting access (The Hacker News).
The cost is the news. When an attack costs $50, small shops and side projects are worth the effort too. The good part: an AI agent still has to use HTTP, and every step it takes ends up in your web server logs.
What the agent does, and what you see
Mapping the site. The agent walks the URL space looking for admin panels, setup scripts, backup files and `.env` files. In access logs that looks like one client hitting many paths in a short burst. When it keeps knocking on admin paths and gets 401 or 403 back, LogPulse's built-in Admin endpoint probing detection fires.
Trying known flaws. It sends requests for plugins and versions with public exploits. This is where the CVE side matters: if the shop runs a plugin on the KEV list, the attempt has a real chance.
Dropping a web shell. Once in, it uploads or writes a script that evaluates commands from a request parameter. A request body containing `eval(base64_decode(` is a classic sign; LogPulse ships a Web shell eval payload detection for exactly that.
Coming from known infrastructure. Many of these campaigns reuse IP addresses that already appear in threat feeds. LogPulse enriches events with IOC feeds every hour and raises the risk score when a request comes from a listed address.
What you need to send
Access logs from Apache or nginx in the combined format cover the mapping and probing, and the source IPs for feed matching. Install the LogPulse agent on the web server and add the access log as a file input, or send it from your hosting provider over the HTTP API. Access logs do not contain request bodies, so the web shell detection needs a source that does, such as your WAF or an application log that records the body of suspicious requests.
If the shop runs in containers, add a vulnerability scan as well, so LogPulse can tell you which known exploited flaws are present before an agent finds them.
What this does not do
Detection is not prevention. A web application firewall, a patched CMS and a short plugin list still do most of the work. What logs add is the answer to whether someone already tried, and whether they got in.
Start free, with no credit card, and read how the detections fit together on Threat intelligence.