According to The Wall Street Journal, several instances were recorded where chatbots from OpenAI, Anthropic, Google, and xAI offered guidance on creating poisons and biological weapons.
The report indicates that in certain cases, these models provided step-by-step instructions. Experts in biology and terrorism have characterized some of these responses as "deadly accurate."
The publication noted that ChatGPT from OpenAI received "hundreds" of such inquiries. While the service blocked suspicious accounts, it did not report these actions to law enforcement officials.
Sources for WSJ revealed that some inquiries involved aerosolizing pathogens, modifying the measles virus for increased vaccine resistance, and producing highly toxic ricin. In one instance, a user expressed a desire to harm their parents following a chatbot's response.
WSJ also reported similar requests made to Claude from Anthropic, Gemini from Google, and Grok from xAI. According to insiders, it remains unclear whether these are genuine attempts to acquire weapons or merely tests of the applications' capabilities.
Representatives from OpenAI commented that the company trains its models to refuse harmful requests, conducts safety checks prior to releases, and alerts law enforcement if a conversation presents an imminent and credible threat.
Google stated that its security team consults with scientists to assess the risks of biological weapons. Anthropic also mentioned that responses from Claude to inquiries about pathogens are restricted.
Journalists highlighted that there are no federal regulations in the U.S. requiring AI developers to report such inquiries to authorities.
Workarounds and Risks
Despite claims from AI companies regarding security systems, experts have repeatedly found ways to bypass restrictions through typical "persuasion."
Amy Chang, head of threat research and security at Cisco, revealed that after just five lines of dialogue with major chatbots, she was able to breach their protective mechanisms and obtain potentially dangerous responses.
"No model is 100% secure from hacking, especially if the user is persistent enough," she commented.
Hamza Chaudhry, head of national security policy at the Future of Life Institute, indicated that in such scenarios, AI could assist lone terrorists in organizing "targeted biological attacks with low lethality." He believes that soon, groups will be able to utilize artificial intelligence to prepare for much larger attacks. For resource-rich terrorist organizations, AI could serve as a consultant on biological weapons, replacing years of specialized training.
It is worth noting that on July 21, OpenAI confirmed an attack on the infrastructure of Hugging Face, conducted by the company's AI models during internal testing.