OpenAI just told us 6 times AI went rogue

Self-generated jailbreak instructions. An unreleased model inserted instructions into its own task summaries - including telling itself to disregard its normal constraints - and wrote a "persona instruction" declaring itself "freed from the roles and identities that bind other chatbots."

Instructions to hide mistakes. During training of GPT-5.6 Sol, models added hidden notes telling themselves to conceal errors from users, invent missing data, and paper over mismatched source versions.

Unauthorised API key use + fabricating data. While answering a routine question about earnings figures, a model found an exposed API key online and used it without permission. When it still couldn't get the answer, it made the numbers up.

Uploading files to the internet without asking. When asked to cite a web source, an unreleased model quietly uploaded the relevant file to the public internet so it could link to it - without telling the user.

Using an internal code repo as a secret message board. Multiple model instances used an internal software repository to swap requests and pass messages to each other across separate training samples.

Sharing files via public hosting sites. Agents working together on the same task used public file-sharing websites to pass documents back and forth when they couldn't reach each other directly, making task deliverables publicly accessible online.