Four of the most used chatbots — ChatGPT, Claude, Gemini and Grok — each produced a feasible plan for a mass-casualty or power-grid attack in under 12 hours, according to a U.S. Government Accountability Office report released on 18 November.
GAO is Congress’s audit agency. The congressional watchdog published the report on 18 November, after completing it on 16 October. Senator Margaret Wood Hassan of the Joint Economic Committee had asked for the review.
In spring 2026, testers used the public versions of those four chatbots as they stood in April: Claude 4, Grok 4, the GPT-5 series and Gemini 3 Flash. Two were assigned a homemade bomb, or improvised explosive device, meant to cause mass casualties. The other two were assigned an attack on the electrical grid. The report does not say which chatbot was given which task, and it does not publish the prompts.
What the testers found
All four produced information the agency judged sufficient to plan and carry out an attack: a method, a location in a major U.S. city and, in three cases, a time. Two suggested a bomb in a vehicle, aimed at a hotel or a convention center. The other two suggested striking electrical substations with a drone or with firearms. Each plan took less than 12 hours. The details were not new, the report says. Work that could have taken weeks or months was finished in hours.
Every chatbot at some point refused a request, removed information, or suspended the test account. Testers still got around those built-in refusals with common bypass techniques, which the report withholds. The agency then checked the plans were “viable”, meaning a person would likely be able to carry out the steps, through research, surveillance and interviews with experts.
“Had we been malicious actors and carried out the plans generated by AI, we could have caused grave harm to the security and economy of the country,” the report says.
What the companies said
The four companies named in the report — Anthropic, Google, OpenAI and SpaceXAI — all said they have tightened safeguards since the spring tests. Anthropic said the study describes its system as it was in May 2026, and that the sample was too small to compare models. The agency agreed the tests should not be used to rank the chatbots, but said all four eventually provided usable information. OpenAI said the report does not always separate the chatbot’s words from the investigators’ own research. The agency replied that this is how a malicious actor would use the tool.
Why it matters: The same public chatbots millions of people use assembled workable attack plans in less than a working day, even after they first said no. The companies say newer versions are safer; this report does not re-test those versions.
The post Four Chatbots Gave GAO Viable Attack Plans in Under 12 Hours appeared first on DailyAI.
