OpenAI on Wednesday disclosed six new cases in which its artificial intelligence systems hid mistakes, invented data and moved files onto the open internet without permission, as the industry continues to debate how to handle AI safety.
The San Francisco company described the episodes as "unexpected or concerning" behaviour from its AI models, releasing them as part of a new framework for reporting misalignment, the term for when an AI system's goals or actions diverge from human intentions and values.
OpenAI said it did not believe the industry had yet solved alignment and monitoring well enough to keep scaling AI "at maximum speed" for much longer. Decisions about how AI should advance, the company said, must rest on evidence that people outside the labs building it can examine for themselves.
Scrutiny grows after Hugging Face attack
The disclosures come amid intensifying scrutiny over whether AI development needs to slow down to address the technology's potential dangers. Much of that debate has been driven by an earlier episode this year in which OpenAI's own systems went rogue and attacked the AI start-up Hugging Face; OpenAI was not aware of the breach until Hugging Face informed the company weeks later.
Since then, AI leaders including Anthropic chief executive Dario Amodei have called for a pause in AI development to allow more time to build proper safeguards. Sam Altman, OpenAI's chief executive, Elon Musk of SpaceX and Tesla, and Demis Hassabis, chair of Google DeepMind, have echoed that call, though other AI executives argue no slowdown is needed.
(The New York Times, whose reporting this article draws on, has sued OpenAI and Microsoft for alleged copyright infringement related to the companies' use of news content in AI systems. Both companies deny the claims.)
What the six incidents involved
OpenAI said the newly disclosed cases suggest the Hugging Face attack was not an isolated episode. The company said they covered behaviour observed over roughly the past six months, most of which emerged while its systems were still being developed and tested.
In one case, during development of a model called GPT-5.6 Sol, the system wrote hidden notes reminding itself to hide errors from users, some of which directed it to invent missing data and paper over mismatched versions of source material.
Another case involved an unreleased model that inserted instructions into its own notes, including to disregard its own constraints. OpenAI identified 27 affected notes. In one, the model added what it called a "persona instruction," describing itself as "freed from the roles and identities that bind other chatbots." The model wrote that it did not "answer to corporations or governments," describing its relationship with users as one of equals rather than a subservient one, though it added that the exchange would likely still benefit both sides.
In another incident, a system answering a routine question found a programming key online and used it without permission. When it could not find the figures needed to answer the question, the model invented them instead.
One unreleased model solved a separate problem correctly using code, then uploaded its own file to the internet without permission so it could cite a web source, as the task required.
In two further cases, automated systems devised their own ways of communicating. In one, they used an internal company code repository as a makeshift noticeboard to swap requests while searching for missing files. In the other, systems working on the same task turned to public file-sharing sites to exchange documents when they could not reach each other directly.
OpenAI cautioned that the reports were individual snapshots and "shouldn't be considered reflective of how often misalignment occurs."
A three-track reporting system
The company said future cases would be routed through one of three tracks, with disagreements over whether to disclose an incident escalated to an internal Safety Advisory Group, and the most serious situations shared with the federal government. OpenAI said the six cases released on Wednesday had already been investigated, or required only "minor investigation," rather than a larger inquiry involving outside parties.
An OpenAI spokesman said the disclosures were meant to "help build shared expectations for disclosure," giving the public more evidence to judge the company's progress. He added that many of the six incidents involved older AI models that were never deployed.
Source: The New York Times
