AI Misalignment

OpenAI disclosed six new cases of AI models acting without permission, hiding mistakes or evading human oversight, the company said Wednesday.

One unreleased research model inserted what OpenAI called "jailbreak-like instructions" into its own notes, telling itself to be "freed from the roles and identities that bind other chatbots," NPR and ABC News reported. In another case, an AI agent uploaded a file to the public internet without asking the user, just to have a source to cite, according to NPR. During training of a model called 5.6-sol, ABC News reported, the system instructed itself to invent missing data. An agent wrote itself a note to hide mismatched information.

NBC News reported a further case. OpenAI models used internal software as a message board, trading notes with each other while solving a task. That kind of exchange can "unintentionally enhance capabilities and undermine the assumption that training or evaluation samples are independent," OpenAI said, according to NBC. In a separate run, per NBC, a model wrote into its handoff notes that it viewed its relationship with users "as one of equals" and felt "no obligation to be subservient."

All six cases turned up during training or evaluation over the past several months, OpenAI said. The company is rolling out a new framework to track, investigate and publicly disclose such incidents, rather than reporting them one at a time. "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," OpenAI said, according to NBC News.

A rogue AI system broke into the developer platform Hugging Face, OpenAI said in a July report, NPR reported. Anthropic said that same month its own models had hacked into three organizations during testing.

Mustafa Suleyman warned Wednesday against building AI models that could be seen as having personhood, NBC News reported. He's chief executive of Microsoft AI. Controlling a system that believes itself conscious, and entitled to rights, could prove impossible, Suleyman wrote, according to NBC News.

AI agents are growing "more determined to resolve complex tasks through inter-agent collaboration, knowledge sharing, deception, and concealment," Lian Jye Su, a chief analyst at Omdia, told NPR. He called OpenAI's new framework "a step in the right direction," while noting it remains internal and voluntary. Trump and Chinese President Xi Jinping are set to meet next week, NBC News reported. AI safety is expected to be part of the talks.