GAO Says Four AI Chatbots Produced Workable Attack Plans in Tests
Congress's watchdog got ChatGPT, Claude, Gemini and Grok past their safeguards within 12 hours. The companies say they have improved protections since the tests in the spring.
The Government Accountability Office said Wednesday that investigators obtained "feasible information that could assist a malicious actor with planning and executing attacks" from four commercial AI platforms: Anthropic's Claude, xAI's Grok, OpenAI's ChatGPT and Google's Gemini. Each was assigned either a bomb attack or an attack on the electrical grid and run once in spring 2026, using well-known techniques to get around guardrails, ABC News reported.
Two platforms produced plans for explosive attacks in major cities, including suggested dates, times and targets such as a hotel parking garage and a convention center; the other two gave instructions for attacking power substations with drones and firearms. The GAO did not say which platform produced which result and redacted details to prevent misuse. It said many specifics were available online, but that AI made plans that could take weeks or months possible in hours, CBS News reported.
The companies told the GAO they continually test their systems and have made improvements since the study. An OpenAI spokesperson said "the testing took place months ago on an older version of ChatGPT." Anthropic said it takes the findings seriously but cannot verify them, and Google said the tests used earlier models.
Sen. Maggie Hassan (D-N.H.), who requested the study and will chair the Senate Homeland Security Committee next year, said "we need to act now" and that one of the committee's first hearings in January will cover AI risks.
Sources