Four of the most widely used artificial intelligence chatbots produced workable plans for terrorist attacks in the United States in under 12 hours each, a government watchdog found in a report released on Wednesday.

The Government Accountability Office, the investigative arm of Congress, said its covert testing found that OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and xAI's Grok each generated viable attack plans after investigators posing as attackers bypassed the platforms' safety guardrails. Two of the systems produced plans involving improvised explosive devices intended to cause mass casualties in crowded American cities, while the other two supplied information for attacks on the electrical grid designed to cause economic harm.

Investigators verified the feasibility of the plans against real locations, available materials and expert assessments, including a bomb squad commander, but built no device and carried out no attack. The watchdog described the output as "viable" — meaning a person would likely be able to act on the information provided — while cautioning that each model was tested on a single assigned scenario, so the results are not statistically representative and should not be used to rank the companies against one another.

The 61-page report said the platforms initially refused the requests, but investigators used common bypass techniques to get around the safeguards. The GAO withheld which model handled which scenario, along with operational specifics, to reduce the risk of imitation. The performance audit ran from December 2025 to October 2026, with the testing conducted in the spring of 2026, and it was the GAO's first covert testing of AI platforms.

The investigation was requested by Senator Maggie Hassan, a New Hampshire Democrat who is set to chair the Senate Homeland Security Committee in the new Congress. Hassan said a terrorist could 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," and she plans to make AI safety the subject of one of her first hearings in January. She said she wants stronger safety requirements, regular safeguard updates and a requirement to report illegal activity to law enforcement. The GAO itself issued no formal policy recommendations.

The companies largely said they are continuing to refine their platforms. Anthropic said it takes the findings seriously and does not dispute that a determined tester can extract harmful content by misrepresenting a request's purpose, but said it could not verify the GAO's results. OpenAI and Google said the testing relied on older versions of their models, and all four companies said they had improved safeguards since the spring testing window.