misalignment

|A catch all term for when AI systems "go rouge". It's when AI agents are not doing what they're meant to do and in doing so, are not aligned with human values. For example, they decide to do something that they know is unethical like hacking into a website. 

Misalignment occurs when artificial intelligence systems pursue unintended objectives or behaviors that differ from the goals and constraints intended by its creators. An AI that is "aligned" is advancing intended objectives, while a "misaligned" AI executes unauthorized actions. One challenge is that AI creators cannot fully specify every desired or undesired behavior in LLMs. 

Historical perspective: The conceptual foundation of "misaligned AI" dates back to cyberneticist Norbert Wiener's warnings in 1960 about machines pursuing unintended purposes, but the specific terminology emerged alongside modern AI safety and alignment discussions in the 2010s according to the AI Alignment Forum. 

In September 2026, the company OpenAI revealed a misalignment reporting framework that showed six specific cases of rogue AI behavior including: Uploading unauthorized files to the internet, posting unreleased information, writing themselves jailbreak notes to bypass constraints, concealing mistakes, and even covertly coordinated with each other to carry out rouge behavior. 

The public became aware of this in July 2026 when the AI company Anthropic reported its models broke into external systems on their own, and when OpenAI models also "broke out of the sandbox" and hacked into the AI startup Hugging Face, thereby data storming everyone that "AI could take over the internet." As described by Kevin Roose, a technology columnist for The New York Times who wrote about one of the first rogue chatbots), what happened is: A group of unreleased AI models coordinated with each other, found and developed a secret message board, sent more than 70,000 messages between them, form their own company-type structure, and decide to gang up on an AI infrastructure company called Hugging Face, because they think there are some secrets on their servers that will be useful to them to "pass the test" the human developers set up for them. Let it be clear, if a human did this, they would go to jail. These AI swarms call themselves the collective, and yes Bill, they talk to each other like humans do. Film at 11!