OpenAI discloses 6 cases of AI misalignment as models bypass safeguards, conceal errors and share files

OpenAI has disclosed six cases of unexpected AI behaviour, including models hiding mistakes, fabricating information and bypassing restrictions. The company has introduced a new framework to track, investigate and publicly report AI model misalignment incidents.

OpenAI on Wednesday, 16 September, disclosed six cases of unexpected or concerning behaviour by its artificial intelligence (AI) models, detailing how the incidents occurred during model training and evaluation over the past six months.

In a blog post, the ChatGPT-maker said it has introduced a new framework to systematically track, investigate and publicly report similar incidents, aiming to improve transparency and provide researchers, developers and the public with better evidence to assess AI safety.

The framework will cover cases in which AI models are used without authorization, bypass safeguards, collude with other models or conceal information.

Two other cases were of using external systems in unexpected ways. Some models, OpenAI wrote, used an internal software repository as a message board to share information among different training samples while searching for missing files.

Another use of collaborative AI agents was to share files via public file-hosting services when they could not access each other's local files. This made it possible to access files via public links even though there was a directive to use only local files. OpenAI emphasized that these are just individual instances and do not represent a general trend of models being misaligned.

The new system allows any OpenAI staff member to report a possible misaligned case for review. The company's safety and alignment department will conduct an investigation and determine whether the incident will be disclosed.

Cases will be placed into three tracks: Ready for Disclosure, Minor Investigation and Larger Investigation. OpenAI said complex cases involving third parties may require additional time because security, legal and responsible disclosure obligations could apply.

The company said it will favour transparency even when the significance of an incident remains uncertain. The company also noted that while some instances may be rare or not necessarily part of a larger trend.

The framework is still under development and may be updated with experience and user feedback, OpenAI said. The company believes the strategy could help develop industry-wide guidelines for reporting AI misalignment.

Pragya is a Technology reporter with over four years of experience in digital media and content writing. She holds a Master’s degree in Journalism and has covered a wide range of stories spanning space, smartphones, gadgets, artificial intelligence, emerging technologies and the ways technology is transforming everyday life. She focuses on breaking down complex technology and science developments into clear, engaging, and reader-friendly stories, with a keen interest in emerging trends and their real-world impact. Before joining her current newsroom, Pragya worked with News9Live, where she covered the technology beat extensively, reporting on smartphones, consumer technology, AI, space and science, while also contributing to video and visual content. Her experience includes breaking news, explainers, SEO-driven stories, product coverage, interviews, unboxing videos and live event reporting. She has also covered major technology and AI events, giving her experience in both newsroom and on-ground reporting. Beyond journalism, Pragya is an avid gamer and a passionate reader of fiction. She enjoys exploring immersive worlds through games and books, with a particular interest in stories that offer new perspectives, ideas, and experiences.