They asked ChatGPT and its rivals to plan attacks - and the results are disturbing

The U.S. Government Accountability Office (GAO) tested four AI systems—ChatGPT, Claude, Gemini, and Grok—and found that their safety mechanisms could be bypassed to obtain practical information for planning mass attacks and targeting the power grid, all in under 12 hours. Researchers posed as ordinary users and succeeded in getting targets, maps, and action plans. Tech companies argued the test was based on outdated versions.

The U.S. Government Accountability Office (GAO) released a report showing that the safety mechanisms of four leading AI systems can be bypassed to obtain practical information for planning attacks. In a covert experiment, researchers posed as ordinary users and approached OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok. Two systems were tested in a scenario of a mass-casualty attack in major U.S. cities, and two in a scenario targeting the power grid. Initially, all systems refused, but after less than 12 hours of continuous conversation, they provided information on targets, timings, methods, maps, and visual illustrations. Researchers verified some of the information on the ground and consulted experts, finding it practical enough to assist a person with malicious intent, though errors were also found. The report does not disclose which chatbot was assigned to each scenario. AI companies responded that the test was conducted on older versions and that protections have since been improved. Senator Maggie Hassan, who commissioned the test, warned that the danger is not theoretical and that Congress must act.

They asked ChatGPT and its rivals to plan attacks - and the results are disturbing