At a glance, the passwords the LLMs created looked secure, much like those that a password generator might spit out. But that’s exactly where the problems arose: Although the AI-generated passwords appeared to be complex and safe to use for securing online accounts, they were actually quite predictable upon closer inspection.
All three LLMs exhibited clearly identifiable patterns in how they created these passwords. These patterns included repeated character strings, predictable password structure, frequent reuse of similar characters, clear biases toward certain numbers and letters, and even duplicate passwords in some cases. Although the AI-generated passwords looked random, they really weren’t. This could easily create a false sense of security if you were to use these predictable passwords for your online accounts.
I’ve seen it personally at work where the AI generates its own metadata files containing uuids that it made up, and they end up being duplicates from elsewhere in the project. Unfortunately I can’t really share links.
I’m sure you could find examples in GitHub issues
Edit: I had an honest look, and can’t seem to find anything that isn’t AI slop in web searches anymore… GUIDs aren’t exactly the most common thing in the first place, so maybe I overstated how common this is.