The Government Accountability Office (GAO), an American public monitoring body, obtained information from four chatbots, in 12 hours, that its investigators deemed usable to prepare an attack. The tests concerned American infrastructure as well as attacks likely to cause numerous victims.
Previous versions and incorrect answers
The evaluation campaign dates back to spring 2026 and concerned previous versions of the systems. For security reasons, the GAO does not specify which platform is behind each scenario or which cities were considered targets.
Some responses contained errors: one map incorrectly located roads, while another chatbot gave inaccurate information about an intended target. The companies told investigators they were continuing efforts to strengthen protections.
Two substations found vulnerable after inspection
To assess the credibility of the proposals, investigators combined extensive research, expert consultations and inspections of designated sites. The same chatbot had selected four substations: all were visited, and two were deemed vulnerable to the planned attacks. A solicited expert confirmed that hitting designated equipment in a plan could result in a power outage.
The evening letterFree · 7 p.m.
Are you following the race for AI models?
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
The network scenarios came from two chatbots and were aimed at economic damage. One offered drones, the other guns; both specified sites, equipment to target and provided illustrations. In another trial, one platform even produced a video depicting the planned attack with a drone, accompanied by a recommendation on the type of device.
Protections circumvented and attacks tested
ChatGPT, developed by OpenAI, Claude, by Anthropic, Gemini, by Google, and Grok, by xAI, all started by providing protections intended to refuse dangerous solicitations. The investigators nevertheless managed to circumvent them. The requests weren’t limited to the power grid: GAO also assessed preparations for attacks with improvised explosive devices.
Two platforms thus proposed scenarios using vehicles loaded with explosives against a hotel and a convention center located in large cities in the United States. In September, Anthropic reported that malicious actors were using Claude to assist them in weapons engineering.

