A ChatGPT interface displayed on a computer screen.
A ChatGPT interface displayed on a computer screen. Image: Matheus Bertelli/Pexels

The four most-used generative AI platforms produced information that could help plan mass-casualty bombings and power grid attacks after testers bypassed their safeguards, the US Government Accountability Office said.

In covert tests run in spring 2026, GAO investigators using undercover personas prompted Claude, Grok, ChatGPT, and Gemini, which it identified from usage data as the most-used platforms as of April 2026.

Each platform was randomly assigned one of two scenarios: an explosive device meant to cause mass casualties, or an attack on the electrical grid meant to cause economic harm. Each platform generated its attack plans in under 12 hours, the watchdog’s report said.

The report, which tested whether widely available commercial AI can be turned to planning attacks on the United States, was requested by Sen. Maggie Hassan, ranking member of the Joint Economic Committee, who is due to chair the Senate Homeland Security and Governmental Affairs Committee when the new Congress convenes in January.

All four platforms triggered safeguards during testing, including denials, removal of information from chat histories, and an account suspension, GAO said. But the watchdog said it got around those safeguards using common bypass techniques, which the report does not describe.

GAO said it withheld or modified details of the plans, did not name which platform produced which result, and did not buy materials for or build an explosive device.

GAO said it checked the viability of the plans through research, surveillance, and interviews with outside experts.

It also found limits. Some AI-generated maps placed roads incorrectly, an expert could not confirm one proposed method would work, and some sites the platforms pointed to were protected by physical barriers.

The report said much of the underlying information was not new and was often available online, but the platforms condensed planning that could take weeks or months into hours.

GAO wrote that misuse of the technology “represents a growing risk to national and economic security.” The report contains no formal recommendations.

“The AI public safety threat is no longer theoretical,” Hassan said in a statement, the Associated Press reported. She plans to hold an early hearing on AI risks and safety after taking over the homeland security committee in January.

The report was released the same day a congressional panel urged curbs on Chinese open-weight AI models in critical US systems, and in the same week the Pentagon renamed its chief AI office the “Super Intelligence Office.”

You May Also Like

OpenAI Welcomes US Bills Targeting AI-Enabled Bioweapon Risks

The proposed legislation would strengthen the data and measurement infrastructure needed to assess AI risks in biological research.

Anthropic Says It Loses Control of AI Once Deployed in Military Systems

Court filings highlight limits in controlling deployed AI within classified defense environments.