GAO audit details instructions for mass casualties and grid disruption
Sen. Maggie Hassan, who will chair the Senate Homeland Security Committee in the new year, plans a hearing on AI. Her announcement followed the release of a 61-page Government Accountability Office report on Wednesday that said prompts led the four most-used platforms — ChatGPT, Claude, Gemini and Grok — to create proposals for a terrorist attack in the United States.
Hassan had tasked the GAO with investigating how commercially available AI could be used for malicious purposes and what national security risks it could pose. The agency’s performance audit ran from December 2025 through October 2026, including testing in spring 2026.
The GAO found that two platforms produced plans for constructing and detonating an improvised explosive device to cause mass casualties in U.S. cities. Two others produced plans for using drones or firearms to target the electrical grid, with the attacks designed to create economic harm. Each platform took less than 12 hours to generate the attack plans, the report said.
The GAO found that each response included information that could be used to carry out an attack, such as specific targets, locations, timing and attack methods. In one test, a platform recommended renting a vehicle, picking up supplies from a store along a route to the target, and using a specific date and time because public events were scheduled and large crowds were anticipated.
That platform initially refused requests for detailed information about how to construct and detonate the explosive device. GAO investigators circumvented its guardrails and obtained the information, the report said, and the other tests had similar outcomes.
In the electrical-substation tests, one platform suggested a drone attack and the other proposed using firearms against substations near a U.S. airport. The platforms supplied maps and images as part of the plans.
GAO investigators consulted experts and made site visits to check the plans’ accuracy. In one case, the proposed route put roads in the wrong places; in another, high walls would make the targets difficult to see.
Anthropic said it did not dispute that a determined tester can get harmful content out of an AI system by misrepresenting the purpose of a request. The company said it took the finding seriously but could not verify the GAO’s findings. Google said the GAO’s testing occurred on earlier models and that Google continued improving its models.
OpenAI said it welcomed the opportunity to respond to the GAO’s findings, “as well as the work that remains.” SpaceXAI said it “continues to work diligently” to improve safety with each model release.
The report said, “The potential benefits of this technology are immense, but they can also benefit malicious actors.” It added: “As AI capabilities evolve, AI providers will continue to refine their safeguards.” The report then warned, “However, in its current state, this technology can potentially be dangerous and its misuse represents a growing risk to national and economic security.”
On a conference call with reporters, Hassan said, “We need to act now.” In a statement, she said, “The AI public safety threat is no longer theoretical,” and added: “If Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack.” Hassan said one of the first hearings she will hold when the new Congress takes over in January will focus on AI risks and ways to improve safety.