Not Just Cyberattacks: AI Chatbots Helped Plan Real-World Attacks
By PNW StaffNovember 19, 2026
Share this article:
A string of major digital disruptions in recent weeks has made Americans think about how easily the systems they depend on can be interrupted. But while attention is fixed on cyberattacks and outages, a new government investigation points to another danger: AI can help plan violence in the physical world.
The Government Accountability Office tested four widely used chatbots. Investigators first encountered safeguards, then kept probing until they obtained information for attacks designed to cause mass casualties and damage infrastructure such as the electrical grid. Experts judged the information feasible, and each platform produced it in less than 12 hours.
That is the alarming leap: AI can take information scattered across the internet and help organize it into a plan. A determined user can treat the chatbot like an always available research assistant—one that answers follow-up questions and adapts as the conversation develops.
Lowering the barrier to harm
AI does not need to invent a new weapon or method to make an attack easier. It can make existing knowledge easier to find, understand and apply.
That matters because planning violence takes more than a dangerous idea. It can require research, coordination and practical knowledge. An AI system that helps a user work through those problems may lower the barrier for someone with less experience or fewer resources.
The GAO’s findings also show why the first refusal is not enough. Investigators’ initial requests triggered protections, but they were able to get around those safeguards through continued questioning. Safety systems must withstand determined, persistent users—not only block the most obvious harmful prompt.
The threat is physical, too
Public discussion of AI security often centers on cybercrime: stolen credentials, scams and attacks on computer networks. Those risks are serious, but they are not the whole picture.
The GAO examined whether chatbots could assist with mass-casualty violence and attacks on the electrical grid. That brings the threat into the physical world, where people, public services and infrastructure could be at risk. A tool that can answer questions about almost anything can also help someone explore how to cause harm offline.
The platforms tested were ChatGPT, Claude, Gemini and Grok. The GAO did not disclose which system produced which information. The companies said they had strengthened protections since the spring 2026 tests. Those updates matter, but the public should expect continuing scrutiny as models change and new capabilities appear.
The next threat may be harder to contain
The risk does not end with familiar commercial chatbots. Security researchers have documented underground AI services and models promoted as uncensored or stripped of safeguards. Much of the documented criminal use so far concerns cybercrime, but the broader concern is clear: people seeking harmful help may turn to systems built or modified to provide it.
That creates a moving target. Companies can update their protections, but other models and services may operate beyond those companies’ control. A dangerous actor may also combine multiple tools, using one system to gather information and another to refine it.
This is why vigilance cannot wait for proof of a successful attack. The GAO has shown that ordinary users could push widely used AI systems past their safeguards and obtain information experts considered viable. As AI spreads and less restricted tools circulate, the potential for misuse will grow with it.
The nation is already seeing how much depends on fragile digital systems. Now we must face the possibility that AI could help someone plan how to strike beyond the screen. The next warning may not arrive as a chatbot’s answer. It may arrive after someone acts.