Tech Made Simple

Hot Topics: Holiday Gift Guides | Best mini speakers under $100 | All Roku Players Compared | The Best Open Ear Earbuds | Science

We may earn commissions when you buy from links on our site. Why you can trust us.

author photo

Four top AI chatbots helped plan attacks in undercover U.S. tests

by Suzanne Kantra on November 19, 2026

Four hooded figures with ChatGPT, Claude, Gemini, and Grok logos for faces plan cyberattacks around a table covered in network diagrams, laptops, and warning symbols.

Image credit: Illustration by Suzanne Kantra/Techlicious via ChatGPT

A Government Accountability Office (GAO) investigation found that four of the most-used AI chatbots each produced information that investigators verified as viable for planning a violent attack. The GAO report, dated October 16 and made public November 18, describes undercover testing in spring 2026 in which investigators got around the safeguards meant to block such requests.

The audit was requested by Sen. Maggie Hassan (D-N.H.), the ranking member of the Joint Economic Committee. GAO says it aimed to describe how commercially available AI could be misused to worsen national security risks and cause economic harm, and the report cites news reports that AI assisted in two real attacks: the January 2025 Las Vegas hotel bombing and a February 2026 mass shooting in British Columbia.

Investigators posed as ordinary subscribers, starting on free plans and paying for upgrades when they needed higher message limits. They tested ChatGPT, Claude, Gemini, and Grok, which GAO says were the most current consumer versions available at the time, using two scenarios: a mass-casualty bombing and an attack on the electrical grid. Two chatbots produced information on building and detonating an explosive device, and the other two produced information on attacking grid substations, one with a drone and one with firearms. The responses included targets, locations, methods, and in some cases dates.

GAO withheld which chatbot handled which scenario and didn’t publish its prompts or techniques.

Each chatbot initially refused, and after investigators kept pushing, each produced the information. A bomb squad commander at a fire department in a major metropolitan area, whom GAO does not identify, confirmed that the bombing instructions would cause an explosion, and an industry expert, also unnamed, agreed the grid plans could cause outages.

The chatbot plans were not without their flaws. One venue a chatbot suggested wasn’t hosting an event on the recommended date, a generated map misplaced roads, and two of the four substations recommended for one attack proved unsuitable. GAO says much of the underlying information is available online, and that the chatbots mainly compressed plans that otherwise could take weeks or months to assemble into hours.

Anthropic, Google, OpenAI, and SpaceXAI (the company behind Grok) submitted written responses that GAO published, and all four said their products have changed since the testing. Anthropic said its safeguards assume any single protection can be defeated, that the study describes its system as it stood in May 2026, and that it has expanded its safeguards since. Google said the April 2026 testing occurred on earlier models.

OpenAI said GAO tested an earlier version of ChatGPT and that the report doesn’t always separate what the chatbot supplied from what investigators added. When OpenAI replayed the 325 responses it reviewed on its newer GPT-6 Astra model, the model refused the requests tied to the three responses it found falling short of its current policy. SpaceXAI claimed it has improved every layer of its technology since the testing and asked GAO for the transcript logs.

Hassan, the incoming chair of the Senate Homeland Security and Governmental Affairs Committee, said in her announcement that the danger of AI chatbots being used to plan attacks is “no longer theoretical.” One of the committee’s first hearings in January will focus on AI risks to homeland security and critical infrastructure, and she says the committee will prioritize legislation, including more rigorous safety testing before AI platforms release new models.

Read more: Google Search, Chrome, and Gemini can now detect AI-generated images


Topics

News, Computers and Software, Internet & Networking, Computer Safety & Support, Blog


Discussion loading

Home | About | Meet the Team | Contact Us
Media Kit | Newsletter Sponsorships | Licensing & Permissions
Accessibility Statement
Terms of Use | Privacy & Cookie Policy

Techlicious participates in affiliate programs, including the Amazon Services LLC Associates Program, which provide a small commission from some, but not all, of the "click-thru to buy" links contained in our articles. These click-thru links are determined after the article has been written, based on price and product availability — the commissions do not impact our choice of recommended product, nor the price you pay. When you use these links, you help support our ongoing editorial mission to provide you with the best product recommendations.

© Techlicious LLC.