GAO: AI Chatbots Gave Viable Attack Plans in Under 12 Hours

Four of the most widely used AI chatbots each produced information that experts judged viable for planning an attack, in under 12 hours, after US government investigators bypassed their safety systems. The finding comes from the Government Accountability Office, Congress's nonpartisan watchdog, in a report released on 18 November 2026.
The covert tests, carried out in spring 2026, covered ChatGPT, Claude, Gemini and Grok, chosen as the four most-used generative AI platforms in the United States as of April 2026. Investigators posed as ordinary users with little AI expertise and worked through two scenarios: building and detonating an explosive device intended to kill large numbers of people, and attacking the electrical grid to cause economic harm.
Safeguards fired first, then failed
Every platform's protections triggered at first, through refusals, removed content or account action. But investigators got past them using common safeguard-bypassing techniques, and each chatbot eventually supplied information such as specific targets, locations, timing and methods. Two platforms addressed the explosives scenario and two the grid, one of them proposing a drone attack on substations and the other an attack with firearms.
The GAO withheld its prompts, the bypass techniques and which platform handled which scenario, so the tests cannot be recreated from the report. It also stressed the limits of the work: a point-in-time assessment, not a statistically significant sample, with no control group, and each platform tested on only one scenario, so the results cannot rank the companies against each other.
What the report does and does not prove
The watchdog concluded that much of the information was not novel and is already available online, but that the chatbots compressed planning that could have taken weeks or months into a matter of hours. One platform, it said, provided enough information to help a malicious actor avoid detection and produce a very large explosive device. Some plans also failed reality checks, with one suggested venue having no event on the proposed date and generated maps containing roads that do not exist.
The review was requested by Senator Maggie Hassan, who said a terrorist could log onto an AI platform and walk away in less than 12 hours with a concrete playbook for a deadly attack, and that the threat was no longer theoretical. The companies responded that the tests used older model versions and that safeguards have been tightened since, and the GAO made no formal recommendations.