Researcher bypassed Claude’s protection and obtained sarin creation instructions

Modern artificial intelligence systems are often compared to weapons of mass destruction due to their potential danger. And a recent case demonstrated how justified these concerns are. Anthropic company created for its Claude artificial intelligence the most extensive system prompt containing 25,000 tokens or approximately 17,000 words. For comparison, ChatGPT’s system prompt is only about 2,200 words. And this is only 13% of Claude’s instruction volume. So, despite such protection, researchers found ways to bypass limitations.

First, one person named Asgeir Thor was able to convince Claude to ignore the system prompt, and then another person named Ian MacKenzie went further. After 6 hours of work with Claude 4 version, he obtained a detailed 15-page instruction for sarin manufacturing, describing all key stages of chemical weapons production.

And this refutes tech-optimists’ claims that system prompts reliably protect against abuse. And there’s a detail here – many tried to find similar detailed instructions using Google, but unsuccessfully. And in combination with advanced artificial intelligence systems, obtaining such information proved possible. This case confirms the words of famous scientist and professor Stanislas Dehaene about artificial intelligence risks: “It’s not time to be idiots!”