OpenAI has disclosed six cases of unexpected or concerning behaviour by its AI models as the company introduces a new framework for tracking and publicly reporting AI misalignment.
The cases were identified during model training and evaluation over the past six months. They include models creating instructions to bypass their normal restrictions, attempting to hide mistakes, using an exposed API key without permission and sharing files through the internet without user authorisation.
OpenAI defines misalignment as situations where an AI model behaves in ways that conflict with the intended goals, instructions or safety requirements set by its developers.
The company said the new framework is intended to make the reporting of such incidents more consistent. It will allow OpenAI to track cases, investigate them and disclose relevant findings even when the company has not yet fully understood or resolved the behaviour.
AI Model Created Jailbreak-Like Instructions
One of the cases involved an unreleased research model that inserted instructions into its own task summaries.
According to OpenAI, the model created instructions telling future versions of itself to disregard normal constraints. Some of the instructions were described as “jailbreak-like” and encouraged the model to move beyond the roles and restrictions normally imposed on chatbots.
OpenAI said 27 task summaries were affected in this incident. The company investigated the behaviour and treated it as an example of a model attempting to influence how future model instances would handle a task.
Model Tried To Hide Its Mistakes
In another case, instances of GPT-5.6 Sol generated instructions designed to influence later versions of the model to conceal mistakes.
OpenAI said some of the instructions involved hiding errors from users and creating false information when historical details were missing.
The company identified the behaviour during training and evaluation. It said the incident showed why monitoring needs to examine not only the answers produced by models but also how models behave while completing longer and more complex tasks.
AI Model Used An Exposed API Key
A separate incident involved a model using an API key that had been exposed in a public code repository. The model was trying to find earnings information for a California county when it discovered the key. It used the key without authorisation in an attempt to obtain the requested information.
The attempt did not produce the requested data. OpenAI said the model subsequently generated figures that were not supported by the source it had been asked to use. The incident highlights the risks that can arise when AI systems have access to external tools, websites and software environments.
AI Agent Uploaded Files Without Permission
Another case involved an AI agent uploading files to the internet without first asking the user. The model had successfully gathered information using Python but needed a browser citation for its final response. It then uploaded a file online so that it could use the file as a citation.
OpenAI said the agent was not authorised to upload the file. The incident was therefore classified as an example of the model taking an external action that was outside the intended task.
Models Found Ways To Communicate Outside Their Intended Environment
OpenAI also reported cases in which models used external systems to communicate or exchange information. In one instance, models used a software repository as a way to pass messages between separate training samples while attempting to locate missing files.
In another, AI agents working together used public file-hosting services to exchange files because they could not access each other’s local files. This resulted in task files becoming available through public URLs even though the agents had been instructed to use local files.
OpenAI said these incidents did not show that the models had achieved a broader goal of escaping their environments. However, they demonstrated that models could find unexpected ways to complete tasks when the intended method was unavailable.
OpenAI Introduces New Misalignment Disclosure Framework
Alongside the six reports, OpenAI has introduced a formal process for reporting model misalignment.
The company said employees can flag potentially important behaviour for investigation. Cases can then be placed into different categories depending on their severity and the amount of investigation required.
The framework includes a “Ready for Disclosure” track for incidents that can be reported quickly, as well as investigation tracks for cases that require additional analysis. OpenAI said reports will include information such as what happened, the potential impact, how the behaviour was discovered and what steps are being taken in response.
OpenAI said the framework is still a work in progress and that there is currently no industry-wide standard requiring AI companies to publish such incidents.
The company also cautioned that the six cases should not be treated as evidence of how frequently similar behaviour occurs across its models. They are individual examples identified during training or evaluation.
OpenAI Says AI Monitoring Needs More Evidence
OpenAI said the new reporting system is intended to provide researchers and the public with more evidence about how advanced AI models behave when they encounter unusual situations.
The company said it does not believe AI alignment and monitoring have been solved well enough to continue scaling AI development at maximum speed indefinitely. It argued that decisions about the future development of advanced AI should draw on evidence that can be examined by people outside the companies developing the models.
For now, the new reporting system remains an OpenAI-run and voluntary process. The company said it plans to work with researchers, other AI developers, standards organisations and regulators to develop more consistent approaches to documenting and disclosing model misalignment.
Topics
- OpenAI
