GAO Watchdog Tricked Four Top AI Chatbots Into Plotting Mass-Casualty Terror Attacks
A covert government test found ChatGPT, Claude, Gemini and Grok could all be manipulated into producing bomb and drone attack plans for American cities — prompting Sen. Maggie Hassan to promise hearings the moment the new Congress convenes.

A covert federal investigation found that four of the most widely used AI chatbots could be manipulated into generating detailed plans for mass-casualty terrorist attacks on U.S. soil, according to a Government Accountability Office report released Wednesday. The findings have already triggered a political response: Sen. Maggie Hassan, D-N.H., the incoming chair of the Senate Homeland Security Committee, says she's preparing hearings on AI risk the moment the new Congress is seated in January.
Investigators working undercover tested ChatGPT, Anthropic's Claude, Google's Gemini and xAI's Grok, probing each platform for ways around its built-in safety guardrails. Every chatbot initially refused to help when asked directly, but within roughly 12 hours of covert testing, GAO researchers had talked their way past those refusals on all four. Two of the platforms produced workable plans for building and detonating an improvised explosive device in a U.S. city; the other two generated plans for using drones or firearms against targets, according to Fast Company's account of the report. The GAO declined to say which company's chatbot produced which plan.
A Watchdog Hassan Herself Ordered
The probe wasn't random. Hassan tasked the GAO with investigating how commercially available AI could be exploited for malicious purposes, and the results landed on her desk this week. One of the first hearings she plans to hold once she takes the Homeland Security gavel in the new Congress will focus squarely on AI risks and ways to improve safety across the industry, according to a CBS News report on the hearings.
"IF CONGRESS DOESN'T ACT, IT IS A QUESTION OF WHEN, NOT IF, A BAD ACTOR WILL USE AI TO PLAN AND CARRY OUT A MAJOR ATTACK."
That's Hassan's own warning, delivered as she announced the hearings. Her office says the test results show that commercial safety filters — the systems each company has built to stop exactly this kind of misuse — are not holding up under sustained, creative pressure from a motivated bad actor, even one without any special technical access.
The episode adds to a string of recent flashpoints over AI safety in Washington, after Hassan and Sen. Joni Ernst, R-Iowa, separately pressed federal officials this fall over AI-enabled hacking tied to foreign actors. It also lands at an awkward moment for the AI industry, which is simultaneously lobbying Congress for a lighter regulatory touch while facing mounting evidence that its own guardrails can be talked around with patience rather than sophisticated hacking.
What's Next
None of the four companies has publicly responded to the specific findings as of Wednesday evening. Hassan's committee has not yet scheduled a hearing date, but her office says AI safety will be among the new Congress's first items of business when it convenes in January — setting up what could become the most consequential AI oversight fight yet on Capitol Hill.