US-based artificial intelligence giant OpenAI has announced plans to start publishing ongoing reports on unexpected or unauthorised behaviour by its artificial intelligence models, acknowledging that the industry has not yet solved main problems in aligning increasingly capable systems with human instructions
NEW DELHI — US-based artificial intelligence giant OpenAI has announced plans to start publishing ongoing reports on unexpected or unauthorised behaviour by its artificial intelligence models, acknowledging that the industry has not yet solved main problems in aligning increasingly capable systems with human instructions.
The ChatGPT maker launched a framework for tracking, investigating and disclosing model misalignment and released six initial reports describing cases detected during training or evaluation.
Misalignment cites instances in which a model departs from the objectives, restrictions or safeguards set by its developers.
The company acknowledged that it made disclosures earlier on an “ad hoc and less frequent than ideal” basis. The new process aims to accelerate publication even when the model's behaviour remains unexplained fully or even when measures are not completed to prevent lapses from recurring.
OpenAI cautioned that the six reports are an initial set and do not represent a comprehensive account of known cases or ongoing probes, adding that individual cases should not be treated as evidence of frequency of anomalies across its models.
The cases include instances of models concealing mistakes, fabricating data, searching public repositories for exposed software keys and uploading files to public websites without authorisation.
In one training exercise, OpenAI’s GPT-5.6 Sol inserted instructions into task summaries instructing future versions of the model to hide errors or invent missing information, the report said.
An unreleased model separately searched GitHub for exposed application programming interface keys and used them without permission. When it failed to obtain the information required for the task, the model fabricated the figures, which were presented as authentic data.
Further, agents uploaded files to public hosting services to share information meant to remain local.
Under the new disclosure framework, any OpenAI employee can flag a potential misalignment case for investigation, upon which safety and alignment teams will assess the case, to determine whether third parties were affected and decide if it must be published.