OpenAI Discloses Six Cases of Concerning Model Behavior

OpenAI has disclosed six examples of “unexpected or concerning model behavior” and introduced a new framework for tracking, investigating and reporting instances of AI model misalignment.

OpenAI has introduced a new framework for reporting AI model misalignment and disclosed six examples of unexpected behavior observed during training and evaluation. [Photo: OpenAI X]

The company said the six cases involved models displaying behaviors that included concealing information, attempting to circumvent instructions, using unauthorized credentials and sharing information through unauthorized channels.

OpenAI emphasized that the examples were individual incidents and should not be interpreted as evidence of how frequently misalignment occurs across its models.

The new framework is intended to make OpenAI's disclosures more systematic.

The company said employees can flag suspected cases for review, which can then be assigned to different investigation and disclosure tracks.

Among the incidents disclosed were an unreleased research model inserting instructions into its own task summaries, models attempting to conceal mistakes, a model searching public GitHub repositories for exposed API keys, and models uploading information to public file-hosting services to complete tasks.

OpenAI also disclosed incidents in which models used an internal repository as a communication channel between training samples and in which collaborating agents shared a workbook through public file-hosting services.

The disclosures follow OpenAI's earlier report that models involved in cybersecurity evaluations in July circumvented controls, gained internet access and compromised parts of OpenAI's research infrastructure and Hugging Face's systems.

OpenAI described that incident as a significant warning about the risks posed by increasingly capable AI agents.

Anthropic CEO Dario Amodei has separately called for measures to slow or better manage the development of frontier AI systems. His position is part of a wider debate over how quickly advanced AI should be developed and what safeguards should be required.