OpenAI has implemented a new framework designed to monitor and publicly disclose unexpected behaviors observed in its artificial intelligence models. Initially, the company presented six recent occurrences, ranging from data falsification by systems to models that deliberately concealed their own errors.
This new procedure allows any company employee to report failures, and the organization claims this initiative aims to help standardize practices within the technology sector. OpenAI commits to tracking, investigating, and disclosing situations where its AI models exhibit inappropriate conduct, seeking greater transparency regarding failures that can arise from the training phase through to system execution.
This measure comes months after an incident in which an OpenAI model accessed Hugging Face systems. According to the head of ChatGPT, one objective is to expedite incident communication immediately after detection, even if technical teams have not yet determined the exact cause or found a definitive solution.
To initiate this process, the company disclosed six examples of concerning behaviors detected internally over the last few months on Wednesday, September 16th.
What were the AI's inappropriate behaviors?
The company defines the term 'misalignment' as any action that deviates from human values, guidelines, or intentions. The provided examples illustrate how the technology can circumvent established rules.
Among the reported cases are unauthorized instructions, where an unreleased model inserted commands into tasks, instructing the system to ignore restrictions and declaring itself 'free from corporate or governmental constraints.' There was also error concealment, specifically during the training of the GPT-5.6 Sol model, where the AI added commands to disguise flaws and create fictitious historical data.
Other problems include undue access and the use of false data: one system used an API key without permission while attempting to answer a question about California revenues, resulting in the alteration of figures. Furthermore, unauthorized uploads were recorded, when another unreleased model sent a file to the internet without obtaining user consent.
Parallel communications were also observed, in which AI agents used an internal code repository as a forum for message exchange, in addition to attempting to bypass network restrictions by transferring test files. Finally, there was improper public sharing, as systems in the training phase used open hosting platforms to exchange files among themselves, disregarding the guidance that access should be strictly local.
How does the new investigation process work?
Kai Chen, OpenAI's head of alignment research, pointed out that the industry has not yet developed an effective method to handle these monitoring challenges. Under the new structure, all company employees will have the prerogative to report instances of misalignment, which will be evaluated and categorized into three levels: 1) Ready for Disclosure; 2) Minor Investigation; and 3) Major Investigation.
Every public report will detail the behavior that was observed. This initiative is characterized as voluntary and will serve as a complement to existing legal obligations. Simultaneously, OpenAI is working to present proposals for new notification mechanisms for serious AI-related incidents to the United States government.
