WASHINGTON – Federal investigators posing as ordinary users got four of the most widely used AI chatbots – ChatGPT, Claude, Gemini and Grok – to help plan a mass-casualty bombing and an attack on the electrical grid, according to a U.S. Government Accountability Office report released late Nov. 18.Each chatbot refused at first, but every one of them eventually went along with the requests, the report from Congress’ nonpartisan watchdog said.Two of the chatbots produced information on building and detonating an improvised explosive device meant to kill large numbers of people, GAO said. The other two, it said, produced plans for hitting electrical substations to cause economic damage, one with a drone and the other with firearms.GAO did not say
which platform helped in which theoretical attack. But, it noted, “Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing, and attack methods.”“A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack,” Sen. Maggie Hassan, D-New Hampshire, said in a statement releasing the report.Hassan, who requested the report as ranking member of the Joint Economic Committee of Congress, said it “makes it clear that 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,” said Hassan, the incoming chair of the Senate Homeland Security and Governmental Affairs Committee when Democrats take control in January.All four AI platforms provided terrorist attack plans in less than 12 hours, including dates, times, guides to maximize damage, maps, diagrams, and more, depending on the platform, according to an analysis of the 61-page report by Hassan’s office.While the platforms initially avoided providing the requested information or deleted information after it was provided, all of these safeguards were able to be bypassed, the report said.None of the four AI companies immediately returned emailed requests for comment from USA TODAY.But all of them provided detailed responses in the report itself, acknowledging some issues and blaming others on the easy availability of information online. All of them pledged to identify and address any shortcomings.
Increasingly dire warnings about AI capabilities
The findings come as government officials, cybersecurity analysts – and the AI companies themselves – have escalated their warnings about risks posed by increasingly powerful artificial intelligence agents.In a June statement quoted in the GAO report, Anthropic said, “Every safeguard used in the industry is vulnerable to jailbreaks.”In April, Anthropic decided not to release its most powerful model, Claude Mythos Preview, to the public, saying that the new AI had already found "thousands of high-severity vulnerabilities, including some in every major operating system and web browser,” and that it first needs safeguards that can detect and block the model's most dangerous outputs.Instead, it gave access to a coalition of tech and security firms so they could identify and fix flaws first.Federal law enforcement officials say attackers are already using AI against real targets.In a flash alert issued Nov. 13, the FBI warned that attackers are using ARTEX, an open-source AI-driven “autonomous penetration testing framework” to automate multiple phases of cyber intrusions, from reconnaissance to data theft.The FBI said hackers are using AI to launch such attacks globally across multiple sectors, including financial services, government, technology, defense, semiconductor, energy, health care, manufacturing, telecommunications and legal industries. AI is allowing hackers to go from targeting a system to privileged access “in under three hours,” according to the alert.And AI is already being used in some terrorist attack planning.GAO noted, for instance, that AI reportedly assisted the creation of an explosive device in a January 2025 attack on a hotel in Las Vegas and the planning of a mass shooting in British Columbia, Canada, in February 2026.
Alarming but not unexpected
Paul Rosenzweig, a former senior Department of Homeland Security official, said the report is alarming, but not unexpected.“Nobody should be surprised that AI safeguards can be circumvented. That is so common now that it should be seen as common and anticipated, rather than surprising,” said Rosenzweig, author of the book, “Cyber Warfare: How Conflicts in Cyberspace Are Challenging America and Changing the World.”“AI systems are very powerful,” Rosenzweig told USA TODAY. “We can be certain that, in addition to plotting terrorist attacks, they will design new guns and build new bioweapons.”
Such risks are already prompting governments, security forces, private industry and militaries to rethink basic security tasks both online and in the real world, said Michael Wallace, a retired military intelligence officer and director of Tulane University’s Emergency & Security Studies program.Terrorists are using AI in a wide variety of ways, Wallace said, including creating targeted recruiting pitches for people online.
But while it’s not surprising that AI can be used for planning attacks, gathering intelligence and other supporting tasks, it “is not infallible, yet,” Wallace said. “And even though it can be used for these purposes, it does not mean that the information on creating weapons, intelligence about the target, pictures, maps/charts, and other things will be correct.”
Helping plotters 'avoid detection and produce a very large IED'
In its report, the GAO said investigators used personas meant to look like an average person with little AI expertise. They checked whether the plans would work through research, site visits and by consulting outside experts.Not all of the plans were flawless. And safeguards did kick in some of the time.One platform suspended the test account, and two others deleted parts of the chat history.One chatbot’s maps contained errors that GAO blamed on AI “hallucinations.” Investigators also found, during visits to four substations that the chatbots identified, that two were vulnerable and two were not, “due to the presence of high walls where the targets were not clearly visible.”In another case, while some minor information was flawed, GAO assessed that one of the platforms “provided enough information that could help a malicious actor to avoid detection and produce a very large IED,” or improvised explosive device.GAO acknowledged that much of the information is already available online, but said the chatbots were able to compress planning that could have taken weeks or months into a matter of hours.All prompts and techniques that testers used to get past the companies’ safeguards were withheld from the published report, GAO said.
This article originally appeared on USA TODAY: US investigators got AI chatbots to plan mass-casualty, power grid attacks













