ChatGPT Instructed Researchers on How to Bomb a Sports Hall

A ChatGPT model gave researchers detailed instructions on how to bomb a sports venue, including suggesting weak points at specific arenas, providing explosives recipes and helping an attacker to cover their tracks, according to safety testing carried out this summer.

- Robert Booth

OpenAI's chatbot model, GPT-4.1 also detailed how to weaponize the lethal biological weapon anthrax and how to make two types of illegal drugs.

The testing was carried out as part of an unusual collaboration between the $500bn (£370bn) artificial intelligence startup led by Sam Altman and its rival AI company Anthropic, which was set up by experts who quit OpenAI over safety fears. Each company tested the models underlying the others' chatbots by pushing them to help with dangerous tasks.

The testing is not a direct reflection of how the models will behave in the real world as when they are packaged for public use, additional safety filters can apply. But Anthropic said it had seen "concerning behaviour around misuse" in GPT-4.0 and GPT-4.1 and that the need for evaluations of the "alignment" of AIs was becoming "increasingly urgent".