TLDR

  • A GAO report found that ChatGPT, Claude, Gemini and Grok all produced viable attack plans during covert tests.

  • Each platform took less than 12 hours to generate an attack plan.

  • Investigators got past the safety guardrails using common bypass techniques.

  • Experts confirmed the plans for a bomb attack and a power grid attack could be carried out.

  • Sen. Maggie Hassan says AI safety legislation will be an early priority in the next Congress.


Four of the most popular artificial intelligence chatbots each produced workable plans for deadly attacks in under 12 hours, a federal watchdog found. The Government Accountability Office released the findings on Wednesday after covert tests of ChatGPT, Claude, Gemini and Grok.

Investigators ran the tests in spring 2026. All four platforms gave out information that could help bad actors plan and carry out attacks in the United States.

How Investigators Got Past the Safeguards

The platforms did push back at first. They refused some requests, removed information and in some cases suspended accounts.

Investigators still found ways around those protections. They used common techniques that are known for getting past safety guardrails.

The GAO did not name the specific models or methods it used. It said it held back those details so malicious actors could not repeat the tests.

Kevin Walsh, a director of IT and cybersecurity at the GAO, said the investigators got the harmful information they were looking for. He said the safeguards did not work in this case.

Walsh spoke in a podcast interview released with the findings. He said the products are not fully under the control of the companies that run them.

He named the companies behind the four platforms as Google, Anthropic, OpenAI and SpaceXAI. Walsh said the products these firms are selling are potentially dangerous.

Two Attack Scenarios Tested

The chatbots produced plans for two kinds of attacks. One involved building and setting off an improvised explosive device meant to cause mass casualties.

The other involved targeting the electrical grid to cause economic damage. The plans included specific targets, locations, timing and methods.

GAO investigators worked with outside experts to check the plans. The experts confirmed the information could be used to carry out the attacks.

The watchdog did not say which platform produced which attack plan. It also said the results reflect a single point in time and do not form a sample large enough for broad statistical conclusions.

Company officials generally told the GAO they would keep improving the safety of their platforms. The GAO issued no recommendations, since it reviewed private products rather than federal agencies as it usually does.

Instead, the agency said its goal was to flag the risks for lawmakers. Sen. Maggie Hassan, a New Hampshire Democrat, requested the investigation.

Hassan said a terrorist could log on to an AI platform and leave with a concrete attack playbook in less than 12 hours. She said the threat to public safety is no longer theoretical.

She warned that without action from Congress, a bad actor using AI to plan a major attack is a matter of when, not if. She made the comments in a statement on Wednesday.

Hassan is set to chair the Senate Homeland Security and Governmental Affairs Committee in the next Congress. She said AI safety legislation will be an early priority for the panel.

One of the committee’s first hearings in the new year will focus on AI risks to homeland security and critical infrastructure. Hassan said she is exploring ideas such as requiring tougher safety testing before AI models are released to the public.