The U.S. Government Accountability Office (GAO) said in a report released Wednesday, Nov. 18 that four leading AI platforms, ChatGPT, Claude, Gemini and Grok, each produced a plan for a terrorist attack inside the United States when investigators asked. Sen. Maggie Hassan, the Democrat who requested the audit, said she will hold a hearing on AI risks and safety when the new Congress convenes in January. The companies generally responded that they are continuing to strengthen their safeguards.
The U.S. Government Accountability Office (GAO) said in a report released Wednesday, Nov. 18 that four of the most widely used artificial intelligence (AI) platforms each produced a plan for a terrorist attack inside the United States when investigators asked. According to CBS News, the platforms are ChatGPT, Claude, Gemini and Grok.
The audit was requested by Sen. Maggie Hassan, a Democrat who will chair the Senate Homeland Security Committee starting next year. Hassan asked the GAO to examine how commercially available AI that anyone can use could be put to malicious purposes and what risks it poses to national security.
All four platforms produced plans in under 12 hours
According to the 61-page report, the audit ran from December 2025 to October 2026, and the testing took place in spring 2026. In tests where investigators concealed their identities, all four platforms produced attack plans in less than 12 hours. Two laid out plans to cause mass casualties with improvised explosive devices (IEDs), and the other two proposed attacks on the power grid using drones or firearms. The GAO said each platform supplied detailed information that could be used in a real attack, including specific targets, locations, timing and methods.
In one test, a platform suggested renting a vehicle and buying supplies at stores along the travel route. It also recommended a specific date and time because a public event expected to draw a large crowd was scheduled. The GAO said the platform initially refused requests for detailed information on building and detonating explosives, but provided it after investigators got around its safeguards.
In the substation attack tests, one platform proposed a drone attack and another proposed a firearms attack on a substation near an airport in the United States. Both plans aimed to cause economic damage, and the platforms also supplied maps and images.
GAO: "Not new information, but work that takes months was done in hours"
Investigators checked the accuracy of the plans through expert consultations and site visits. In one case, a road was marked in the wrong place; in another, they found that a high fence made the target hard to see.
The GAO assessed that the information the AI produced was not new and that much of it is available on the internet. But it noted that planning that could take weeks or months was finished in a matter of hours with AI's help, and that AI can suggest ideas and methods that would be hard to come up with through basic research. "The potential benefits of this technology are enormous, but it can also benefit malicious actors," the GAO said, adding that in its current state it can be dangerous and, if misused, poses a growing risk to national and economic security.
Companies: "We continue to strengthen safeguards"
The report also includes responses from Anthropic, Google, OpenAI and SpaceXAI. In a letter to the GAO, Anthropic said it does not dispute that a persistent tester who misrepresents the purpose of a request can draw harmful content out of AI, and that it takes the findings seriously. However, it said it cannot independently verify the report's results. Google said the GAO tested earlier models and that it continues to improve its models. OpenAI said it welcomes the chance to respond to the GAO's findings and the remaining challenges. SpaceXAI said it continues to work to improve safety with each model it releases.
"We must act now," Hassan told reporters on a conference call. She said that when the new Congress convenes in January, one of its first hearings will address AI risks and ways to improve safety. In a statement, she said, "The threat AI poses to public safety is no longer theoretical," adding, "If Congress does not step in, it is not a question of whether a malicious actor will use AI to plan and carry out a large-scale attack, but when."

