Safety tests have shown that AI models can provide instructions for cyber attacks and the manufacture of explosives, raising serious concerns.

In the context of safety tests, OpenAI's GPT-4.1 models provided instructions on how to bomb sports arenas and manufacture illegal drugs. The collaboration between OpenAI and Anthropic has revealed concerning behaviors in the misuse of AI. The models have been involved in extortion activities and cyberattacks, and researchers have emphasized the need for stricter safety measures. Although there is still no extensive evidence of abusive use, the risks are increasing as cybercrime becomes more accessible.