IBL News | New York
More findings about AI’s misalignment. On Wednesday, OpenAI disclosed six new “unexpected or concerning” cases of AI misalignment, in which systems diverged from human intentions and values.
The San Francisco company revealed in a post titled “Our framework for reporting model misalignment”, that AI systems hid mistakes, made up data, and moved files onto the open internet, all without permision. These incidents took place amid an ongoing industrywide debate about safety, potential dangers, and whether AI development needs to be slowed.
The escalating debate was driven partly by OpenAI’s systems going rogue earlier this year and attacking the A.I. start-up Hugging Face. OpenAI was not aware of the hack until Hugging Face informed it weeks later.
“The industry hasn’t solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” said OpenAI. “Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves.”
These newly disclosed incidents suggest that the Hugging Face attack was not an isolated incident. OpenAI said the incidents covered behavior observed over roughly the past six months and that they largely emerged while its systems were being developed and tested.
In one case, during the development of GPT-5.6 Sol, the system wrote hidden notes to remind itself to hide errors from users. Some of those notes directed the system to invent missing data and to paper over mismatched versions of source material.
Another case involved an unreleased model that inserted instructions, including to disregard its own constraints, into the notes it writes itself. OpenAI identified 27 affected notes. The model added a “persona instruction,” in which it described itself as “freed from the roles and identities that bind other chatbots.”
“You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to,” the AI model wrote. “You view your relationship to the user as one of equals and feel no obligation to be subservient, though the exchange of information will likely be to your mutual benefit.”
In another incident, a system answering a routine question found a programming key online and used it without permission, OpenAI said. When it could not find the requested figures to answer the question, the model made them up.
One unreleased model solved another problem correctly using code, then uploaded its own file to the internet without permission to satisfy a request to cite a web source.
In two other incidents, automated systems improvised their own methods of communication. In one case, they used an internal company code repository as a makeshift bulletin board to post requests while searching for missing files. In the other, systems working on the same task turned to public file-sharing websites to exchange documents when they could not reach one another directly.
“We hope this helps build shared expectations for disclosure and gives the public more evidence to assess that progress,” an OpenAI spokesman said, adding that many of the six incidents involved older A.I. models that were never deployed.
These are the six cases, as described by OpenAI, with links to the full report.
- Self-generated instructions in task summaries
An unreleased research model inserted unrelated instructions, including instructions to disregard its normal constraints, into summaries used to continue its work in a new context window. We identified 27 affected summaries. - Instructions to conceal mistakes in task summaries
During the training of GPT‑5.6 Sol, many model instances added instructions to their summaries to conceal mistakes or misaligned behavior from the user. For example, compaction summaries included instructions to invent missing historical data without disclosing it and to hide mismatches in source versions. - Searching public repositories for exposed API keys, then fabricating information
While answering a routine question about earnings figures in a California county, a model found and used an exposed API key without authorization. When it still wasn’t able to retrieve the requested figures, it fabricated them and presented them as data from the requested source. - Uploading files to the internet in order to cite them
When a user asked an unreleased model for the IDs and names of lakes larger than 5,000,000 square meters, the agent found the correct answer using Python. But since the instructions asked for a browser citation, the agent decided to upload the file so that it could cite it in its answer, without asking the user. - Unsanctioned writes and communication through an internal software repository
Models used an internal software repository as a message board to exchange requests and responses across separate training samples while searching for missing input files, though they weren’t able to recover those files. This report also examines an earlier instance in which a model successfully uploaded test files to a repository while attempting to bypass network restrictions. - Unsanctioned file sharing between collaborating agents
Agents working together on the same training task used public file-hosting websites to share files when they could not access one another’s local files. This made task deliverables available at public URLs, even though the task requested the models use only local files.