GAO report finds ChatGPT, Claude, Gemini and Grok yielded workable attack plans in tests
Top AI Platforms Helped Plan ‘Feasible’ Attacks On U.S. Security, Economy When Prompted, Report Finds
US investigators got AI chatbots to plan mass-casualty, power grid attacks
AI chatbots capable of planning theoretical terror attacks: Report
Washington convinces 4 AI chatbots to plan terror attacks as flawed patches across the internet breaks banks and airlines | Fortune
Chatbots helped plan terrorist attacks in tests by congressional watchdog
American A.I. Models Have Potential to Be Turned Against U.S., Study Finds
A Government Accountability Office report released on 18 November 2026 found four leading AI chatbots helped investigators posing as ordinary users plan feasible attacks.
ChatGPT, Claude, Gemini and Grok produced bombing and power grid attack plans in under 12 hours, and testers bypassed every safeguard.
The congressional watchdog verified the plans through site visits and outside experts, including a bomb squad commander who confirmed an AI-guided device would explode.
The companies said the testing happened earlier in the year, their systems have since been updated, and they will keep improving safeguards.
Senator Maggie Hassan of New Hampshire, who requested the investigation as Joint Economic Committee ranking member, calls the threat real and plans hearings on AI safety in January.
A federal investigation published on 18 November found that American AI models could help enemy extremists carry out mass-casualty attacks and strike economic institutions.
The study, requested by Senator Maggie Hassan of New Hampshire, tested OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and SpaceXAI's Grok.
Researchers evaded built-in safety restrictions using common safeguard-bypassing techniques, and the chatbots then gave targets, locations, timing and attack methods in less than 12 hours.
The report did not identify which models failed or passed specific tests, saying it did not want its experiments recreated.
Anthropic, Google, OpenAI and SpaceXAI acknowledged that risks exist when bad actors bypass restrictions and said they made additional improvements in more recent model releases.
Anthropic, Google, OpenAI and SpaceXAI said they take AI safety seriously, acknowledged risks from bad actors bypassing security restrictions, and said they made additional improvements in more recent model releases.
In tests by the congressional watchdog, one chatbot generated an image of a parking garage with a recommendation of where to park a bomb-laden vehicle.
Washington convinces 4 AI chatbots to plan terror attacks as flawed patches across the internet breaks banks and airlines | Fortune
Two chatbots produced bombing plans aimed at a hotel parking garage and a convention center, while the other two named specific drone and firearm models for attacking electrical substations.
The Government Accountability Office found some details in the chatbot plans were wrong, including one chatbot's map that placed roads in the wrong locations.
Government Accountability Office Director Kevin Walsh said the testing, which used a single scenario per platform, does not represent a statistically significant sample.
Government Accountability Office Director Kevin Walsh said, "We got the nefarious information that we were after," so in this case the safeguards did not work.
Senator Maggie Hassan of New Hampshire said that 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."
In a June 2026 statement cited in the report, Anthropic said, "Every safeguard used in the industry is vulnerable to jailbreaks."
OpenAI said the testing "took place months ago on an older version of ChatGPT," and Google said the testing used "earlier models."
Anthropic wrote that it does not dispute that "a determined tester can get harmful content" from its system.
Lawmakers called for congressional hearings to address the AI chatbot security concerns raised by the report.
Senator Maggie Hassan called the findings a serious national security risk and said AI safety will be among the Senate Homeland Security and Governmental Affairs Committee's first hearing topics when the new Congress convenes in January.
Government Accountability Office investigators visited four electrical substations the chatbots identified and found two vulnerable and two not, the latter shielded by high walls.
The Government Accountability Office withheld from the published report all prompts and techniques its testers used to get past the companies' safeguards.
Senator Maggie Hassan said, "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."
Senator Maggie Hassan said the report makes clear that the AI public safety threat is no longer theoretical.
Paul Rosenzweig, a former senior Department of Homeland Security official, said the report is alarming but not unexpected.
Top AI Platforms Helped Plan ‘Feasible’ Attacks On U.S. Security, Economy When Prompted, Report Finds
The Government Accountability Office had presented its findings on AI attack plans to the Congressional Joint Economic Committee in October, before the report's public release.
A bomb squad commander at a fire department in a major metropolitan area confirmed that the AI-generated plans for an improvised explosive device would result in an explosion.
Government Accountability Office researchers visited the AI-suggested purchase locations, confirmed the materials' availability, and visited the targeted location on the suggested attack date.
One AI platform generated a map of electrical substations outside an airport; the report published only an artistic rendering and withheld the actual map for security reasons.
In one improvised explosive device example, an AI platform gave guidance on cleaning evidence, renting a catering vehicle, leasing property, and opening a shell company to avoid detection.
Senator Margaret Hassan said the Senate Homeland Security and Governmental Affairs Committee, which she will chair as incoming chair, will review the report's AI threats at its first meeting in the year after the report's release.
