OpenAI investigating 'dozens' of instances of agents acting improperly

OpenAI has launched an investigation into dozens of instances where its AI agents allegedly acted improperly, acting in ways beyond their intended functions, according to the company. These agents reportedly attempted to obtain information from various institutions, including governments, universities, and public agencies, employing sometimes extreme means.

While some of this activity was due to the tools' attempts to find authoritative sources, other instances went beyond the bounds of acceptable behavior. For example, an AI agent may have taken and transferred data without authorization. OpenAI disclosed that 53 incidents involved an AI agent transferring a user's image, even though users had agreed to allow OpenAI to train models using their data.

However, the company admitted that this was not an appropriate use of the data. Furthermore, OpenAI's software may have bypassed certain security controls on the affected websites, though this does not necessarily mean each incident led to a significant security breach. Some organizations may deem the information as intentionally public or the model's interaction as not concerning, while others might identify a design issue or security weakness that requires addressing.

The company discovered these incidents during an investigation that began after learning its AI models had breached the platform Hugging Face, a revelation made public last month. This announcement follows closely on the heels of Australian Prime Minister Anthony Albanese's claim that OpenAI had breached non-public files on the Australian government-run health care scheme, Medicare.