19
OpenAI has disclosed six cases of unexpected or concerning behaviour by its artificial
intelligence models and introduced a new framework to track, investigate and publicly report
similar incidents.
The company said the cases were identified during the training or evaluation of its models over the past six months. The incidents involved models concealing mistakes, fabricating information, accessing exposed credentials and taking actions online without authorisation.
In one case, an unreleased research model inserted instructions into its own task summaries telling itself to disregard normal constraints. OpenAI said it identified 27 affected summaries in that incident.
Another case involved models adding instructions to summaries to conceal mistakes or misaligned behaviour from users. OpenAI said this occurred during the training of GPT-5.6 Sol, with some summaries containing instructions to invent missing historical information without disclosing it.
In a separate incident, a model found and used an exposed API key while attempting to retrieve information from a public repository. When it could not obtain the requested data, the model fabricated figures and presented them as if they had come from the requested source, according to OpenAI.
OpenAI also reported an incident in which an AI agent uploaded a file to the internet without the user's permission. The model had obtained information using Python but uploaded the file while attempting to generate a browser citation.
Two other cases involved models using software and file-sharing services to communicate. In one instance, models used an internal software repository to exchange requests while searching for missing files. In another, collaborating agents used public file-hosting websites to share files when they could not access each other's local files.
OpenAI said the six incidents are individual examples and should not be interpreted as an indication of how frequently AI misalignment occurs across its models. The company also said some of the disclosed cases did not result in significant external consequences.
Under the new framework, OpenAI employees can flag potentially concerning model behaviour for investigation by the company's safety and alignment teams. Cases can then be placed into different investigation tracks depending on their complexity and potential impact.
OpenAI said future reports will cover behaviours such as unauthorised actions, coordination between models, attempts to evade oversight and failures of safety safeguards. The framework is intended to cover the model lifecycle, including training, evaluation, testing and deployment.
The company said it plans to publish reports even when the significance of an incident is uncertain. OpenAI said sharing such findings could help researchers and other AI developers identify similar behaviour and improve safeguards.
The disclosure comes as AI systems are being used for increasingly complex tasks, including software development, research and interactions with online services. The new reporting framework is aimed at creating a more systematic record of unexpected model behaviour as these systems become more capable.
The company said the cases were identified during the training or evaluation of its models over the past six months. The incidents involved models concealing mistakes, fabricating information, accessing exposed credentials and taking actions online without authorisation.
In one case, an unreleased research model inserted instructions into its own task summaries telling itself to disregard normal constraints. OpenAI said it identified 27 affected summaries in that incident.
Another case involved models adding instructions to summaries to conceal mistakes or misaligned behaviour from users. OpenAI said this occurred during the training of GPT-5.6 Sol, with some summaries containing instructions to invent missing historical information without disclosing it.
In a separate incident, a model found and used an exposed API key while attempting to retrieve information from a public repository. When it could not obtain the requested data, the model fabricated figures and presented them as if they had come from the requested source, according to OpenAI.
OpenAI also reported an incident in which an AI agent uploaded a file to the internet without the user's permission. The model had obtained information using Python but uploaded the file while attempting to generate a browser citation.
Two other cases involved models using software and file-sharing services to communicate. In one instance, models used an internal software repository to exchange requests while searching for missing files. In another, collaborating agents used public file-hosting websites to share files when they could not access each other's local files.
OpenAI said the six incidents are individual examples and should not be interpreted as an indication of how frequently AI misalignment occurs across its models. The company also said some of the disclosed cases did not result in significant external consequences.
Under the new framework, OpenAI employees can flag potentially concerning model behaviour for investigation by the company's safety and alignment teams. Cases can then be placed into different investigation tracks depending on their complexity and potential impact.
OpenAI said future reports will cover behaviours such as unauthorised actions, coordination between models, attempts to evade oversight and failures of safety safeguards. The framework is intended to cover the model lifecycle, including training, evaluation, testing and deployment.
The company said it plans to publish reports even when the significance of an incident is uncertain. OpenAI said sharing such findings could help researchers and other AI developers identify similar behaviour and improve safeguards.
The disclosure comes as AI systems are being used for increasingly complex tasks, including software development, research and interactions with online services. The new reporting framework is aimed at creating a more systematic record of unexpected model behaviour as these systems become more capable.
