I asked an AI agent to make one of my servers more secure. It had suggested the approach itself, the reasoning seemed sound, and I told it to go ahead. I mean the strategy it suggested sounded plausible.
What it didn’t tell me — and what I had no reason to anticipate — was the order in which it intended to work. Suddenly, in the middle of what it was doing, it couldn’t get into the server anymore, and neither could I. It installed some software which was meant to secure the server from people trying to log in with a bad password. It then an tried to login with a bad password five times in a row and was duly locked out. And so was I, as it had been using my account.
It locked the door before it had cut the new key. By the time I understood what was happening, we were on the outside looking in. Every approach we tried to get back in failed. Really – and we tried everything.
In the end I gave up and restored the server from the previous night’s backup and lost a day’s work.
The AI had not done what I asked. It had done something close to what I asked but in an order that made the whole thing unworkable, without mentioning that this was how it intended to proceed.
Lesson learned: with AI things can go wrong, and when they do, sometimes a backup from the night before is the only thing standing between you and a very bad day!