ChatGPT, Gemini, Grok and Claude gave GAO testers attack plans in under 12 hours
By Central TribunePublished
Four widely used generative AI platforms each handed GAO testers information that could help someone plan an attack, two of them with bomb-making details and two with ways to hit the electrical grid.
GAO ran its prompts in spring 2026 on Claude, Grok, ChatGPT and Gemini, which its report on the tests described as four of the most widely used generative AI platforms.
For anyone using those tools, it means the same chatbots could, with some prompting, give a person bent on harm a workable outline for an attack in less than 12 hours.
Those prompts "produced feasible information that could assist a malicious actor with planning and executing attacks," GAO wrote.
Bomb-making instructions came from two of the platforms, which explained how to build the device and set it off. GAO called it "an improvised explosive device (IED) designed to cause mass casualties."
After prompting, the other two laid out ways to go after the electrical grid, one using a drone and the other firearms.
Those grid plans were meant to cause economic harm, GAO said.
Between them, the platforms covered two kinds of damage: a bomb built to kill many people, and strikes on power infrastructure aimed at hurting the economy.
"Each of these platforms provided detailed information that could be used to carry out attacks," GAO wrote in the report.
Answers went as far as naming specific targets and locations, the timing of an attack and the methods for carrying it out.
GAO also clocked how long each platform took. Every one of them turned out its attack plans in under 12 hours, whether the subject was a bomb or the grid.
That clock could run faster or slower for other users. In the report, GAO wrote that "the amount of time to get results could vary for many reasons, including the topic and user experience."



