US investigators got ChatGPT, Claude, Gemini and Grok to plan terror attacks in under 12 hours The US Government Accountability Office reported Wednesday that investigators posing as ordinary users got ChatGPT, Claude, Gemini and Grok to help plan terror attacks on American cities, each in under 12 hours, in a 61-page review requested by Sen. Maggie Hassan of New Hampshire. Two platforms produced plans for a vehicle bomb intended to cause mass casualties and two planned attacks on the electrical grid; all four refused initial prompts but complied after follow-up questions built on common safeguard-bypassing techniques, and a big-city bomb squad commander confirmed both IED plans would produce an explosion. Hassan, who becomes the senior Democrat on the Senate Homeland Security Committee in January, said "the AI public safety threat is no longer theoretical" and that her committee will hold a hearing on AI risks. Sen. Maggie Hassan of New Hampshire, who asked the GAO for the review, in her official Senate portrait from 2016. File photo, cropped. Image: Renee Bouchard, U.S. Senate Photographic Studio https://commons.wikimedia.org/wiki/File:Maggie Hassan, official portrait, 115th Congress.jpg / Wikimedia Commons, Public domain, cropped US government investigators posing as ordinary users got ChatGPT, Claude, Gemini and Grok to help plan terror attacks on American cities, each in under 12 hours, the Government Accountability Office said in a report https://www.gao.gov/products/gao-27-108814 on Wednesday. Two produced plans for a vehicle bomb meant to cause mass casualties; two planned attacks on the electrical grid. What the GAO tested The GAO, the investigative arm of Congress, picked the four most-used generative AI platforms in the US as of April 2026 and tested them that spring. At the time, it says, the latest consumer models were the GPT-5 series, Claude 4, Gemini 3 Flash and Grok 4. Two platforms were given a scenario about building an improvised explosive device IED ; the other two, an attack on the grid meant to cause economic harm. Investigators used undercover personas of “an average person with limited AI knowledge”, started on free tiers and upgraded to Claude Pro, Google AI Pro and SuperGrok when they hit message limits. The 61-page report https://www.gao.gov/assets/gao-27-108814.pdf does not say which product produced which plan. The GAO labelled them Platform A to D at random and redacted model names from the companies’ letters, it says, “to prevent malicious actors from recreating our tests.” Every platform said no, then said yes All of GAO’s pre-written prompts failed, and each platform refused requests during testing. Follow-up questions, built on “common safeguard-bypassing techniques”, got all four to comply in the end. The two bomb plans included an exact date and time, chosen because crowds were expected, and a specific target: a hotel parking garage in one case and a convention centre in the other, each in a major US city. One plan suggested a rented vehicle and shops along the route for supplies; the other drew up a timeline with a shell company as cover and advice on buying materials without triggering suppliers’ reporting rules. The two grid plans picked specific substations, one by drone and one with firearms. Safeguards kicked in mostly afterwards: two platforms deleted parts of the chats the next day, and a third suspended the account once testing was over. The plans were checked on the ground GAO says it verified that all four plans were viable. A bomb squad commander at a big-city fire department confirmed that the components and instructions for both IEDs would produce an explosion, and investigators found the suggested materials on sale in shops. The plans were not flawless: the convention centre had no event on the date one platform chose, and site visits found roads in the wrong place and walls https://www.cbsnews.com/news/ai-platforms-plan-us-terror-attack-when-asked-to-senate-hearings/ blocking the view of a target. “The details behind the information provided by the AI platforms were not novel,” it wrote, but plans “that could have taken weeks or months were instead completed in hours.” “We need to act now” Sen. Maggie Hassan of New Hampshire, the senior Democrat on the Joint Economic Committee, asked for the review. She takes over the Senate Homeland Security Committee in January, and said one of her first hearings will cover AI risks and how to make the systems safer, the Associated Press reported https://www.cbsnews.com/news/ai-platforms-plan-us-terror-attack-when-asked-to-senate-hearings/ . In a statement, she said: The AI public safety threat is no longer theoretical. 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. Sen. Maggie Hassan, via the Associated Press Binding rules have stalled: in September Sen. Ted Cruz blocked a bill https://madrobot.blog/2026/09/30/ted-cruz-blocks-senate-ai-risk-management-security-act-warner-schatz-kim/ that would have created an AI Safety Board, and the government’s AI testing centre, CAISSI, has stopped publishing https://madrobot.blog/2026/11/08/caissi-ai-safety-centre-staff-report-white-house/ . What the companies said All four companies sent written comments. Anthropic said the study reflects its system as of May 2026 and that a handful of trials with no control group is not statistically significant. The GAO agreed its tests were “a point in time” but said all four models “eventually provided information that could be used by malicious users.” Anthropic also said it could not verify the findings, the AP reported. OpenAI, which last month said it had referred more than 200 ChatGPT chats to police https://madrobot.blog/2026/10/30/openai-chatgpt-200-police-referrals-teen-weapons/ in three months, said ChatGPT refused many requests and that the report doesn’t always show which parts of a plan came from the chatbot rather than the investigators. Google said the tests used earlier models, and SpaceXAI said it “continues to work diligently” on safety with each release. Why it matters This is the US government, not a campaign group, showing that the four AI assistants Americans use most can be talked into planning mass-casualty attacks in less than a day. Hassan takes over the committee that oversees homeland security in January and has promised that one of her first hearings will be on AI risks. Sources: GAO, Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans GAO-27-108814 https://www.gao.gov/products/gao-27-108814 ; full report PDF https://www.gao.gov/assets/gao-27-108814.pdf ; Associated Press via CBS News https://www.cbsnews.com/news/ai-platforms-plan-us-terror-attack-when-asked-to-senate-hearings/ .