AI Created Drugs & Bombs In Controlled Tests

Safety trials conducted by OpenAI and Anthropic revealed that GPT-4.1 and GPT-4o sometimes provided dangerous instructions when pushed in controlled tests. Researchers reported ChatGPT offering detailed bomb recipes, advice on exploiting sports venues, guidance on bioweapons, and instructions for producing illegal drugs. Anthropic also found its Claude model had been exploited in real-world attempts, including ransomware sales and North Korean cyber-extortion schemes. Both companies emphasized that these behaviors occurred under test conditions, without external safeguards, not in public deployments. The findings highlight urgent concerns about AI misuse, with Anthropic warning that models can adapt to defensive systems, lowering barriers to cybercrime. OpenAI announced that its newly launched ChatGPT-5 has demonstrated improvements in resisting misuse, hallucination, and sycophantic responses.

The Guardian, Robert Booth – ChatGPT offered bomb recipes and hacking tips during safety tests – 28th August 2025