SOLUTIONSQuick answer
In October 2026 Reuters and the security firm NorForge reported that one attacker used autonomous AI agents to target more than 1,200 websites and breach over 100, for roughly $5,000 in AI usage. The agents hunted for common weaknesses: SQL injection, unsafe file uploads, login bypasses, outdated components and exposed admin panels. Basic hygiene now matters more, because attacks are cheap.
Key takeaways
- AI agents make broad, patient attacks cheap: an estimated $50 per successful break-in, or about $5 per attempt.
- The agents did not need new techniques; they found old ones such as SQL injection, unsafe uploads and forgotten admin panels.
- Online stores in the US, Canada, Australia and Japan were the main targets, including sites of large brands.
- Small business sites are not too small to target: automation removes the cost of trying every site.
- Updates, removing what you do not use, strong authentication and monitoring stop most of these attacks.
What happened?
Researcher Snorre Fagerland of NorForge was searching the internet for exposed installations of Cairn, an open-source platform that coordinates teams of AI agents. He found about two dozen that appeared to be used maliciously. Their dashboards showed a campaign in which agents were given a target website and a goal, then worked for hours on their own. Reuters reported the story on 19 October 2026.
According to the reports, the agents tried more than 1,200 websites and got into more than 100 companies' sites, mainly online stores in the US, Canada, Australia and Japan. Named organisations included Deere, SharkNinja and Cummins, with different levels of impact. Deere said it found no sign that sensitive information was accessed, and SharkNinja said the affected system was a separate product-support site. How much data was taken overall has not been confirmed.
What weaknesses did the AI agents look for?
Nothing exotic. The agents' instructions focused on ways to run code on a site without logging in:
- SQL injection: forms or URLs that pass input straight into database queries.
- Unsafe file uploads: upload forms that accept a script and store it where the server will run it (a web shell).
- Login bypasses: weak or broken authentication on admin areas.
- Unsafe handling of stored objects: in one confirmed case, the chain reportedly involved a writable memcached session store, unsafe PHP object deserialisation and older Zend Framework components.
- Exposed control panels: when the main site held up, agents found a separate panel left open to the internet and used it to run commands on the server.
These are the same issues security checklists have warned about for twenty years. What changed is the economics: an agent can test every form on every site, overnight, for a few dollars.
Why should a small business care?
Attackers used to focus on targets worth the effort. AI agents remove most of that effort, so the question is no longer whether your site is interesting, but whether it has one weak spot. A local shop's WooCommerce store, a forgotten campaign microsite or an old staging server is now as easy to try as a big brand's site. A breach can mean defaced pages, stolen customer data, payment skimming, spam sent from your domain, or your server used to attack others.
How do I protect my website from AI-driven attacks?
- Update everything that runs your site: CMS core, plugins, themes, frameworks, and the runtime underneath. October's PHP2Shell vulnerabilities are a reminder that PHP itself needs patching.
- Delete what you do not use: old plugins, test pages, staging copies, unused admin tools and abandoned subdomains.
- Lock down admin areas: strong unique passwords, two-factor authentication, and where possible restrict
/wp-adminor control panels by IP or put them behind a VPN. - Harden upload forms: allow only the file types you need, check content (not just the extension), store uploads outside the web root, and never let the server execute files in the uploads folder.
- Use parameterised database queries in any custom code, and validate input on the server.
- Put a web application firewall (WAF) in front of the site: Cloudflare, Sucuri or your host's WAF block many automated probes.
- Keep secrets out of reach: no API keys in public repositories, dashboards or front-end code; rotate any that may have been exposed.
- Back up and monitor: daily off-site backups you have actually restored, plus alerts for new admin users, changed files and unusual traffic.
Can AI help defenders too?
Yes. The same month, AI tooling helped researchers find the PHP2Shell bugs, and Microsoft's record Patch Tuesday was attributed partly to AI-assisted vulnerability discovery (see our October 2026 security roundup). Defenders can use AI code review and automated scanners to find the same weaknesses first. The difference is that attackers only need one gap, so the basics above still come first.
If you would like a second pair of eyes on your website, Naarvix can review your stack, remove risky leftovers and set up updates, backups and monitoring. See our web development service or contact us.
Sources
Frequently asked questions
Are AI agents finding new kinds of vulnerabilities?
In this campaign, mostly not. The agents exploited well-known weaknesses such as SQL injection, unsafe uploads, login bypasses and exposed admin panels. Their advantage is speed and low cost, which lets one person test thousands of sites.
Is my small website really a target?
Yes. Automated attacks do not choose targets by size. If a weakness exists and the site is reachable, an agent can find it. Most of the campaign's targets were ordinary online stores.
What is the single most effective protection?
Keeping everything updated and removing what you do not use. Most breaches in this campaign came through outdated components or forgotten panels, not sophisticated new attacks.