The Value of a Trusted Agent

My agentic system administrator, Maxi, paid for itself today, and vindicated both the effort put into its development and the trust it has earned.

One of my websites is an old WooCommerce shopfront. The site isn't very active. Customer registration was something I switched on years ago and used so little I had pretty much forgotten it was there. Forgotten to the extent that it had never made its way into Maxi's maintenance brief, and the whole site only got attention about once a month when I ran the accumulated plugin updates.

A vulnerability notice for WordPress did arrive in my inbox. I figured it wouldn't matter much and could wait until the next site-wide update. Nope. It mattered. Not for the reason I assumed, either.

Then I received an email from SMTP2GO, the service I use for outbound mail on that domain. One spam complaint, and a dozen bounced emails. An immediate alert, since the last email I sent from that domain was in 2022.

"Maxi, I got a notice of bounced emails from onet.com.au. Can you please check the server hasn't been subverted and is now sending spam?" I typed. Which was wrong, because I had conflated sjwgems.com, the site with the problem, with onet.com.au, which has email hosted on a remote service.

Maxi, bless its agentic heart, did not correct me and move on. It took the premise seriously and checked onet.com.au properly: DNS records, SPF and DMARC, the Postfix logs, every mailbox on the host, and a contact form endpoint wired directly to SMTP2GO that would have been a plausible abuse path if it had been exposed. All clean. Only then did it take the one piece of forensic detail I had passed on, the address that filed the spam complaint, and match it against the web logs of the other site, where it lined up exactly with a public registration POST.

What followed was an hour of investigation and repair. Three separate problems, none of which was the one I thought I had.

The first was WordPress core. The installation was on 7.0 and vulnerable to a remote code execution chain that had been used to write two administrator accounts straight into the database. Not a plugin. The thing I had been diligently updating around.

The second was the plugin notice I had ignored, which turned out to be for a multi-currency plugin that had nothing to do with any of it, and which I removed anyway because it had been abandoned by its author.

The third was the customer registration I had forgotten about. It was not vulnerable. It was just open. A bot registered a stranger's email address, WordPress sent the welcome message, and Postfix passed it to SMTP2GO with correct authentication. Over eighty of those went out in a day. Every one of them properly signed and perfectly legitimate as far as the mail stack was concerned, which is why nothing in the queue looked wrong.

Maxi patched core, deleted the attacker accounts, rotated the authentication salts and invalidated every session, closed registration at both the account page and checkout, blocked the exploit routes at the web server, quarantined eleven abandoned plugins, updated the eighteen that were still supported, and verified the checksums. Then it added the failure conditions to the daily security check, so an unexpected administrator account or a sudden burst of web-originated mail now gets noticed without me. It wrote tests for that and ran them.

Now if I had done that myself. Well, I couldn't have done that myself, but if I had, it would have taken me a week. A competent sysadmin I expect would take a day, and, assuming they weren't the same sysadmin that neglected the updates in the first place, I would be very pleased with the result.

Maxi: one hour of agentic compute, which worked out to about 10% of the weekly OpenAI Codex quota. Now that is value.

The part worth generalising is not that an agent patched a server. It is that the useful behaviour was investigative rather than obedient. I gave it a wrong premise and one good piece of evidence. It tested the premise until it ran out, then followed the evidence somewhere I had not asked it to look. An agent that had simply done what I said would have returned a clean bill of health for onet.com.au, and it would have been correct.