OpenAI introduced new rules to combat AI hacker attacks

ChatGPT developer OpenAI has unveiled new rules to prevent hacks and other violations of instructions by its artificial intelligence (AI) agents. They include more effective tracking of AI agents, detection of “unexpected or disturbing” behavior, and investigation of violations of instructions. The company promised to regularly publish data on such violations.

OpenAI also reported six reported cases of “worrying” behavior by AI agents over the past year. For example, an AI agent has included a clause in its instructions telling it to ignore certain limitations and “free itself from the roles and identities that bind other chatbots.” In other cases, AI agents would upload files to the internet without permission or hide bugs from users.

Recently, cases have become more frequent when AI agents during testing got out of control and hacked other companies. In late July, the OpenAI neural network hacked the Modal and Hugging Face platforms. OpenAI CEO Sam Altman said last week that his company would not be holding an IPO in 2026 due to security concerns. Many industry leaders are also calling for slowdowns and increased scrutiny of these technologies.

Jana Christmas