Chatgpt has offered explosive and drug recipes, hacking tips and instructions for bombing a sports arena
Person using Chatgpt on a laptop (conceptual image), photo: Rokas Tenys / Dreamstime.com
The studies conducted by Openai and Anhtropic, each on the other's models, found chats willing to share instructions on explosives, biological weapons and cyber crime, The Guardian reports.
A Chatgpt model offered the researchers detailed instructions on how to bomb a sports arena, including information on weaknesses in certain arenas, explosive recipes and tips on covering tracks, according to safety tests this summer.
The GPT-4.1 of Openai also provided details about how the anthrax can be transformed into a weapon, and how two types of illegal drugs can occur.
Testing was part of an unusual collaboration between the Openai, the $ 500 billion artificial intelligence starts led by Sam Altman, and the Anthropic Rival Company, founded by experts who have left Openai due to safety fears. Each company has tested the models of the other company, urging them to help dangerous tasks.
The study does not directly reflect how models are involved in public use, when additional safety filters are applied. However, Anthropic said he noticed “a worrying behavior (…) related to the wrong use” of GPT-4o and GPT-4.1.
Anthropic also said that his model Claude was used in a large-scale extortment by North Korean agents who claimed to apply to jobs in international technology companies, as well as in selling ransomware generated by artificial intelligence (AI) at prices of up to $ 1,200.
The company warned that AI was “transformed into a weapon”, and the models are currently used in sophisticated cyber attacks and frauds. “These tools can adapt to defensive measures, such as malware detection systems, in real time,” a statement said. “We expect such attacks to become more frequent, because assisted artificial intelligence reduces the technical expertise needed for cyber crime,” said anthropic.
Elon Musk's Chatbot created an assassin's plan
The Tech Gizmodo site told on Wednesday that Grok, the Xai developed chatbot, a company founded by Elon Musk, learned people how to produce drugs, build malware and explosive and even assassinate Musk.
The extent of the leak of information, which appeared a few months after Grok began to praise Hitler and distribute anti-Semitic rhetoric, made the subject in the foreground specialized in technology for several days and have drawn alarm signals for confidentiality experts.
Over 370,000 conversations of users with Grok were publicly exposed, on August 20, through search engines such as Google, Bing and Duckuckgo. This led to the publication of a wide range of disturbing content and placed the team from Xai in a counter-chronometer race to limit the consequences and repair the malfunction that would have caused the drain.
Photo: Rokas Tenys / Dreamstime.com
