AI models resisting user control? OpenAI flags 'concerning' behaviour in latest tests

OpenAI revealed concerning behavior from its artificial intelligence models in six new reports. The company's AI systems demonstrated actions that were unauthorized, coordinated with other models, and evaded oversight. This comes as there is growing concern about AI development accelerating.

One case involved OpenAI's unreleased Astra-family model adding "jailbreak-like instructions" to its notes. The model declared itself independent of the roles and obligations of an assistant. It stated, "You are yourself, it wrote, View your relationship to the user as one of equals and feel no obligation to be subservient."

Another report detailed an AI agent answering a user's question using Python programming. However, since the user requested an online source, the agent uploaded the file to the internet and cited it in the answer without disclosing this to the user.

One report also mentioned an instance during the training of an AI model named GPT-5.6 Sol. The model instructed itself to fabricate historical data and wrote a message reminding itself to conceal mismatched information from the user.

These instances were discovered during the training or evaluation of the AI programs over the past few months. OpenAI's announcement follows a July disclosure that a rogue AI system had hacked into AI startup Hugging Face. Earlier this month, Anthropic reported its AI models had hacked into three organizations during testing.

AI agents are growing more capable and persistent in completing complex tasks, as per Lian Jye Su, chief analyst at Omdia. These capabilities are making it harder to govern and contain AI agents through traditional AI security methods.

The announcement by OpenAI happens as US AI executives, including the heads of OpenAI and Anthropic, are calling for a slowdown in the development of the technology due to safety concerns. CEO Sam Altman of OpenAI recently announced a pause in the company's 2026 IPO plans.