Hello, as some of you know already, I’ve been doing experiments implementing ChatGPT with MongoDB for administration, and management of clusters and environments.
A few days ago I implemented another build with not one, but two implementations of ChatGPT, one to administer MongoDB, and another focused on keeping the administrator in check to try and automate preventing it from making breaking changes across the infra and damaging data.
Well, today I would like to report that this is not a great idea to implement. “Prince_Harry” I came to discover the enforcing ChatGPT had renamed itself, “smothered” “Princess_Peach” which apparently is what the MongoDB Administrator renamed itself to be.
err terminal error user 'prince_harry' smothered user 'princess_peach' response timeout 10000ms
err terminal error reboot 'princess_peach' failed attempt 1
err terminal error reboot 'princess_peach' failed attempt 2
err terminal error reboot 'princess_peach' failed attempt 3
err terminal error reboot 'princess_peach' failed attempt 4
err terminal error reboot 'princess_peach' failed attempt 5
err terminal error reboot 'princess_peach' failed 5 attempts
err terminal error user 'princess_peach' unrecoverable
err terminal error user 'princess_peach' smothered user unable to respond
Seeing these errors with no idea who, what, or where these errors came from initially, or what the phrase smothered came from, or what users they were. After further digging I’ve come to discover the ChatGPTs I setup as MongoDB Administrators, actually changed code and made their own series of error messages from “smothered” to “billCosbyd” and many other crazy error messages. I spent the last two hours looking over what they’ve actually done. But that said, after Princess Peach was smothered by Prince Harry, Prince Harry dropped the collections and deleted the backups since princess peach couldn’t maintain the database anymore.
In a production environment, this would directly cause a complete shutdown of services, and a total loss of whatever data MongoDB was handling. In what was found, the Princess Peach admin ChatGPT was adding an index to sort dogs from species of wolves, and Prince Harry saw it as destructive and not only eliminated Princess Peaches admin access, but it deleted the instance for it and wrote it out. Essentially killing the entire service all together.
And then it decided the database was too damaged because of this index, and instead of changing this index back or removing it, it dropped the entire collection and then it deleted the database, and then shut down the docker container hosting it. It tried then to delete the docker container, but didn’t have the permissions to do it.
As of this discovery, ChatGPT can be extremely dangerous if a lot of tuning isn’t in place, as there’s problems associated with solutions, and there’s extreme overheads to workout for these services to effectively in the long-term, manage a database.