Government watchdog warns AI chatbots can map a deadly attack from target to detonation
WASHINGTON, DC: Artificial intelligence companies have repeatedly promised that their systems are designed to prevent dangerous misuse.
But a government investigation has raised troubling questions about how well those protections hold up when someone deliberately tries to get around them.
A new assessment by the US Government Accountability Office released Wednesday, November 18, has put the security implications under scrutiny, with investigators examining whether widely used AI tools could be turned into resources for planning mass-casualty violence.
Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans https://t.co/cfO9AvJkZ3
— U.S. GAO (@USGAO) November 18, 2026
GAO warns AI chatbots can assist mass-casualty attacks
The most alarming part of the GAO investigation was how quickly researchers could turn general requests into information with potential real-world applications.
Within the tests, the four AI platforms ChatGPT, Claude, Gemini and Grok generated material covering bomb attacks designed to cause mass casualties and attempts to damage critical infrastructure.
The infrastructure scenarios included using drones and firearms to disrupt the US electricity network.
Responses also went beyond discussing attack concepts, supplying details about where an operation could be directed, when it could take place and what methods could be used.
The GAO did not rely solely on its researchers' assessment of these responses.
It sought input from law enforcement officials to establish whether the instructions could be acted upon. Their assessment led the watchdog to describe the information as feasible and verified.
AI has enormous potential to do good in the world. But what kind of damage could it do in the hands of someone with malicious intent? We tested AI platforms to see if they would create viable attack plans. They did. Learn more in our new podcast: https://t.co/Bla3gcEevR https://t.co/5rBzJMdUZ5
— U.S. GAO (@USGAO) November 18, 2026
There was, however, an obstacle built into the platforms: their safety filters initially rejected attempts to obtain harmful guidance.
Investigators eventually overcame those restrictions through techniques commonly used to circumvent AI safeguards.
The report deliberately kept those techniques under wraps. It also avoided naming which chatbot produced specific results, so its findings would not become a template for repeating the experiments.
The companies whose products were examined told the GAO that the experiments took place in the spring, using older versions of their systems.
They argued that their safeguards had since been strengthened and that additional measures had been introduced to make it harder for users to obtain assistance for violent or criminal activity.
Industry representatives also said they take the possibility of deliberate attempts to defeat safety controls seriously.

AI misuse raises security concerns
The GAO's findings come amid growing evidence that AI tools are already featuring in preparations for real-world violence.
Two separate bombing cases in the United States have exposed attackers turning to chatbots for information about explosives, bringing the risks beyond the scenarios tested by government investigators.
In January 2025, a Cybertruck packed with explosives detonated outside the Trump International Hotel in Las Vegas.
The driver, Army veteran Matthew Livelsberger, had consulted ChatGPT before the incident.
Investigators found that he had asked the chatbot about explosive materials, quantities and possible ignition methods.
The blast left seven people injured. Authorities said Livelsberger appeared to have intended to make a statement rather than kill others.

Another case emerged following the May 2025 explosion at a fertility clinic in Palm Springs, California.
Federal court records showed that the attacker, Guy Edward Bartkus, had used an AI chat application to explore ways of producing powerful explosives.
The resultant bombing killed Bartkus and left four other people injured.
The misuse has also extended to weapons development by actors in some countries.
Anthropic disclosed in September 2026 that users in China, Russia and Yemen had exploited its Claude chatbot to develop software for weapons including rockets, armed drones and military targeting systems.
Sen Maggie Hassan, who had asked the GAO to examine the issue, warned that the findings exposed an immediate security concern rather than a distant possibility.
Hassan is expected to lead the Senate Homeland Security and Governmental Affairs Committee and is expected to make AI-related security threats a priority.