Not entirely a mistake, but ...
Just as "security" had begun sneaking into the minds of IT people - 1998 - I was working for a very young start-up doing security testing. We had a bunch of tools to perform various tests trying to wriggle information from systems, send mails via systems that shouldn't, work around firewall rules (if there were any) etc.
And a few tools to perform destructive tests like shutting down systems. We didn't use those unless the client specifically asked us to do so and authorized it - in writing.
One day I was working on-site at a customer with a large dinosaur of an IBM system, trying to make my way in. The sysadmins were quite smug about this "security test", and I was supposed to run the full set of tests - including the potentially destructive ones. So late in the afternoon I dig into that section of the toolbox and begin poking around the network interfaces using SNMP. The server joyfully provides all sorts of interesting info - the configuration, IP-adresses of systems it is connected to etc. Okay, let's see what else is possible - we had been authorized to try the potentially destructive tests, so I fired off the SNMP "write" command with the default password to switch off the primary network interface.
Which it did.
Quite a bit of frantic activity followed to get the system back online. I just leaned back and pulled out the (virtual) popcorn.
Another big-iron experience - mid 1990's - was when connecting one of those newfangled "unix" systems at a branch office to the REAL mainframe computer via an X.25 connection. All went well until we should try sending some data: The mainframe crashed hard enough that a full IPL (reboot) was needed. Turned out there was a bug in the mainframe X.25 comms stack which the unix system accidentally triggered.