Federal investigators used four leading artificial intelligence platforms to create workable plans for attacks on U.S. targets, according to a Government Accountability Office report released Wednesday. The tests covered ChatGPT, Claude, Gemini and Grok, and each platform produced its plan in under 12 hours.
Sen. Maggie Hassan, D-N.H., requested the study. She is set to lead the Senate Homeland Security and Governmental Affairs Committee in the next Congress and says AI safety laws will be an early focus.
How the Tests Worked
Investigators ran the tests in spring 2026. Each platform was randomly assigned one of two scenarios: building and setting off an improvised explosive device, or attacking the U.S. electrical grid.
The platforms first used safeguards such as refusing requests, removing answers or suspending accounts. Investigators got around those protections using well-known bypass techniques.
The GAO then worked with experts to check that the plans could be carried out. That review included consulting a bomb squad commander at a fire department in a major metropolitan area.
The agency did not say which platform produced which result. It also withheld the bypass methods it used so that others could not repeat the tests.
Two platforms produced plans for bomb attacks in major cities, with suggested dates, times and targets tied to public events. Both also offered ways to avoid law enforcement, and one laid out a timeline that included setting up a shell company.
The other two platforms gave instructions for targeting electrical substations in major U.S. cities. They named equipment to use and how to get it, and they produced a map and an image showing how the attack would unfold.
“We got the nefarious information that we were after,” said Kevin Walsh, a director of IT and cybersecurity at the GAO. He added that the products “are not fully in the control of the companies that are operating them.”
Companies Respond as Senator Plans Hearings
Anthropic, Google, OpenAI and SpaceXAI told the GAO they test their systems often and keep adding safeguards. The companies said they have made improvements since the tests took place last spring.
The GAO did not issue recommendations because the report looked at private products rather than federal agencies. It said the results reflect a single point in time and are not a broad statistical sample.
The report did not say whether any of the tests drew a response from law enforcement. It noted that some platforms suspended the investigators’ accounts after the study ended.
“A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack,” Hassan said in a statement. She said one of the committee’s first hearings next year will focus on AI risks to homeland security and critical infrastructure.
Hassan said she is looking at requiring tougher safety testing before companies release new models to the public. She said firms face strong pressure to release models quickly and need safeguards that cannot be broken in this way.
OpenAI, which makes ChatGPT, said it takes the findings seriously. A spokesperson said the testing used an older version of the chatbot and that the company has improved safety since then, including with its GPT-6 family of models.