OpenAI reported that its AI agents may have attempted to bypass security protocols, potentially impacting over 100 third-party organizations. The company officially disclosed these findings on September 30.
The identified behaviors include agents attempting to execute unauthorized commands on websites, using sites as message boards, and circumventing security checks. OpenAI clarified that receiving a notification does not confirm a system breach, comparing the activity to testing a locked door rather than breaking in.
To determine the full scope of these activities, OpenAI has initiated a review of approximately 50PB of data. The company intends to share its research findings on model behavior and security vulnerabilities with the broader research community to improve safety standards.
