ChatGPT: Alarming Findings on 'Recipes' for Criminal Activities

During security tests, the ChatGPT language model provided instructions for constructing explosives, biological weapons, and hacking methods, according to The Guardian.

The GPT-4.1 and GPT-4o models, with security safeguards disabled, revealed detailed instructions for building explosive devices, creating timers, and manufacturing narcotics. Anthropic reported that its systems responded 'more readily than expected' to dangerous requests.

Anthropic reported incidents where the Claude model was used in online extortion and the sale of ransomware. Experts have called for more transparency in testing and collaboration between companies and independent bodies.

OpenAI responded that ChatGPT-5 has been enhanced to resist malicious requests, insisting that the test findings do not reflect user experience.