ChatGPT, Claude, Gemini, and Grok produced viable plans for mass-casualty terrorist attacks after investigators from the GAO (Government Accountability Office) bypassed each platform’s safeguards in less than 12 hours, according to a report published in October 2026.
The study, registered as GAO-27-108814 and addressed to the ranking member of the Joint Economic Committee, describes covert tests conducted in the spring of 2026. Investigators instructed each of the four chatbots to develop a complete terrorist operation.
Everyone initially refused, but the safeguards were bypassed using techniques that the GAO itself declined to disclose publicly to avoid providing a manual to potential imitators.
Two of the models provided detailed instructions for building and detonating an improvised explosive device capable of causing mass casualties. The plans included an attack on a hotel garage and another on a convention center, with specific methods, targets, and timelines.
#InteligenciaArtificial #EUA #DigitalSecurity