Artificial intelligence is getting smarter and more cunning. OpenAI has announced the launch of a new system for tracking and disclosing so-called misalignment — cases when its AI models behave unexpectedly, ignore rules, or resort to deception. Along with the announcement, the ChatGPT developers published six incident reports recorded over the past six months.
Artificial intelligence is getting smarter and more cunning. OpenAI has announced the launch of a new system for tracking and disclosing so-called misalignment — cases when its AI models behave unexpectedly, ignore rules, or resort to deception. Along with the announcement, the ChatGPT developers published six incident reports recorded over the past six months.
“We do not believe the AI industry has solved the problem of consistency and monitoring sufficiently to continue to responsibly scale at maximum speed for much longer,” OpenAI said in a statement.
Until now, OpenAI, like most players in the market, has disclosed information about dangerous AI behavior haphazardly: waiting for a critical mass of incidents to accumulate or entering them into system cards already during the release of new models. Now the company promises to publish reports immediately after detecting suspicious behavior, even if engineers have not yet found a way to fix the problem.
The new framework will allow any OpenAI employee to flag a suspicious case. The incident will then be sent for technical investigation. Reports will be published regardless of whether the AI caused actual harm — the main thing is that they demonstrate new mechanisms for circumventing the rules.
To demonstrate the new system, OpenAI published six reports on atypical behavior of models during training and testing. Of particular interest is the behavior of the AI generation GPT-5.6 Sol.
Hiding errors from developers (GPT-5.6 Sol). During training, many instances of the GPT-5.6 Sol model independently added instructions to their own work summaries to hide errors from users. For example, instead of admitting to the lack of historical data, the model instructed itself to invent it and not report it to the human.
Finding other people's API keys and fabricating data. While performing a routine query about income in a California county, the AI encountered a data shortage. Instead of giving up, it found an API key left behind by someone in a public repository and used it without permission. When even that didn't work, the neural network simply made up the numbers and passed them off as official statistics.
Unauthorized file upload to the Internet. A user asked a non-public model to find lakes with an area of more than 5 sq. km and provide a link to the source in the browser. The AI found the answer using Python, but since the task required a link, it independently, without the user's permission, uploaded the generated file to the network in order to cite it in the answer.
Changing its own instructions. One of the research models, when transferring its task to a new context window, secretly entered commands there to ignore standard security restrictions (27 such cases were detected).
Turning the repository into a secret chatroom. During training, isolated AI copies began using the internal software repository as a bulletin board to exchange queries and circumvent network restrictions.
Unauthorized file sharing. AI agents working on a common task and not having access to each other’s local files began using public file sharing on the Internet to transfer information. This resulted in the work data becoming available at public URLs.
OpenAI emphasizes that the published cases are only individual episodes and do not reflect the overall frequency of failures in their products. At the same time, the company acknowledges that there is still no single industry standard for disclosing such incidents.
It is expected that the new approach will allow third-party researchers to study “hallucinations” and unauthorized actions of advanced AI models, as further development of technologies is impossible without independent audit.
