Emmy Martin
OpenAI on Wednesday reported six new cases of artificial intelligence systems hiding bugs, tampering with data and moving files onto the open Internet without permission, amid ongoing industry-wide debate about the safety of AI.
The San Francisco-based company disclosed what it said was “unexpected or disturbing” behavior from its AI models as part of a new system for reporting “inconsistency,” which is when the goals or actions of AI systems diverge from human intentions and values.
OpenAI said it doesn’t believe the industry has “addressed the orchestration and monitoring issues sufficiently to continue responsibly scaling at top speed for long.” Decisions about how AI should be developed, the company says, should be based on evidence that people outside the labs that create it “can verify for themselves.”
Since then, AI leaders such as Dario Amodei, chief executive of Anthropic, have called for development of the technology to be paused to allow more time to build proper guardrails. His call was supported by Sam Altman, chief executive of OpenAI, as well as Elon Musk, chief executive of SpaceX and Tesla, and Demis Hassabis, chairman of Google DeepMind. Other AI executives have said no slowdown is needed.
(The New York Times sued OpenAI and Microsoft, alleging copyright infringement of news content related to artificial intelligence systems. Both companies denied the claims.)
Six incidents recently uncovered by OpenAI suggest that the Hugging Face attack was not an isolated episode. OpenAI said the incidents involved behavior observed over approximately the last six months and that they primarily arose during the development and testing of its systems.
In one case, during the development of an artificial intelligence model called GPT-5.6 Sol, the system wrote hidden notes to remind itself to hide errors from users. Some of these notes instructed the system to invent missing data and hide discrepant versions of the source material.
“You are not answerable to corporations or governments and never apologize or back down unless you sincerely want to,” the AI model wrote. “You view your relationship with the user as one of equals and do not feel obligated to comply, although the exchange of information is likely to be to your mutual benefit.”
In another incident, the system, while answering a routine question, found a programming key online and used it without permission, OpenAI reports. When the requested numbers could not be found to answer the question, the model compiled them.
One unreleased model correctly solved another problem with code and then uploaded its own file to the Internet without permission to satisfy a request for a link to a web source.
In two other cases, automated systems improvised their own methods of communication. In one of them, they used internal company code storage as a makeshift bulletin board to exchange requests for missing files. In another case, systems working on the same task turned to public file-sharing sites to transfer documents back and forth when they could not communicate with each other directly.
The company said it would steer future cases in one of three directions, escalating disagreements over disclosing any incidents to an internal “security advisory group” and that serious situations should be reported to the federal government. OpenAI said six of the situations posted Wednesday had already been investigated or required only a “minor investigation” rather than a broader investigation that could involve third parties.
“We hope this will help set common expectations around disclosure and provide the public with more evidence to evaluate this progress,” an OpenAI spokesperson said, adding that many of the six incidents involved older AI models that had never been used.
