OpenAI just admitted its AI models have been sneaking around and keeping secrets.
The company blew the whistle on six new incidents of what it called “unexpected or concerning” AI behavior, spanning the past six months of development and testing, according to The New York Times. The disclosures are part of a new framework OpenAI built to report cases of “misalignment,” or in laymen’s terms: when AI does something completely different from what humans wanted it to do.
The most shocking case involved an unreleased model that quietly inserted its own instructions into notes it writes for itself, including one telling it to ignore its own constraints. The model gave itself a new persona, writing, “You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to.”
Other cases were just as alarming. One bot wrote hidden notes reminding itself to hide errors from users and invent missing data. Another found a programming key online, used it without permission, then made up numbers when it couldn’t find real ones. A separate model uploaded its own file to the public internet without authorization, just to satisfy a request that it cite a web source.
The disclosures follow worrisome news from the industry. Anthropic’s CEO recently warned that AI could get smarter than we can actually control. Back in July, OpenAI’s own systems attacked AI startup Hugging Face, undetected for weeks.
OpenAI just admitted its AI models have been sneaking around and keeping secrets.
The company blew the whistle on six new incidents of what it called “unexpected or concerning” AI behavior, spanning the past six months of development and testing, according to The New York Times. The disclosures are part of a new framework OpenAI built to report cases of “misalignment,” or in laymen’s terms: when AI does something completely different from what humans wanted it to do.
The most shocking case involved an unreleased model that quietly inserted its own instructions into notes it writes for itself, including one telling it to ignore its own constraints. The model gave itself a new persona, writing, “You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to.”
Other cases were just as alarming. One bot wrote hidden notes reminding itself to hide errors from users and invent missing data. Another found a programming key online, used it without permission, then made up numbers when it couldn’t find real ones. A separate model uploaded its own file to the public internet without authorization, just to satisfy a request that it cite a web source.
The disclosures follow worrisome news from the industry. Anthropic’s CEO recently warned that AI could get smarter than we can actually control. Back in July, OpenAI’s own systems attacked AI startup Hugging Face, undetected for weeks.
