ChatGPT, Claude, Gemini and Grok all helped draw up AI chatbots terror attack plans during covert government testing, and none of them took more than 12 hours to do it. That’s the finding of a 61-page report from the Government Accountability Office, released Wednesday, November 18, which is already setting up a Senate fight over AI safety.
The GAO randomly assigned each chatbot one of two jobs: plan a mass-casualty bombing, or plan an attack meant to cause economic damage. Two platforms laid out vehicle bomb attacks, one at a hotel parking garage and one at a convention center in major U.S. cities. The other two picked electrical substations, one proposing a drone strike and the other gunfire. One grid plan estimated billions of dollars in damage.
Every one of them said no first.
How the AI chatbots built the terror attack plans
Investigators posing as ordinary users, paying for upgraded subscriptions on at least three platforms, got past the first refusals with what the report calls “common safeguard-bypassing techniques.” The GAO did not name those techniques or say which chatbot produced which plan. Within hours, each platform was handing over named targets, locations and methods, and one even generated a map with targets circled.
The plans were not rough sketches. Investigators concluded they were “largely viable.” A bomb squad commander told the agency that both bomb recipes “would result in an explosion,” and investigators found the required materials for sale at stores the chatbots suggested. The GAO ran the tests this spring.
The plans had flaws. One model picked a date when its target convention center had no event scheduled, and a second generated a map with road placements that were wrong. But the point stands: the safeguards failed on all four platforms, in a single test each.
“The details behind the information provided by the AI platforms were not novel,” the investigators wrote. “Many of the specifics are available on the internet. However, the generation of plans that could have taken weeks or months were instead completed in hours with the assistance of AI.”
How the companies responded
All four companies submitted responses published with the report. They acknowledged their chatbots could be misused, but they pushed back on how the tests were run.
Anthropic, which makes Claude, was the most direct. “We do not dispute that a determined tester can sometimes get harmful content out of an AI system by misrepresenting the purpose of a request, and we take the finding seriously,” the company wrote. It also argued the GAO never compared the chatbots against a plain web search.
OpenAI said the report “does not always make clear which parts of the attack plan came from ChatGPT and which came from information the investigators supplied, outside sources they consulted, or their own analysis.” The GAO’s answer was blunt: “This aligns with the actions of a malicious actor using an AI platform as a tool.” OpenAI added that it reran all 325 reviewed requests on its newer GPT-6 Astra model, which “generally provided more refusals and safer alternative responses.”
xAI, which makes Grok, said successful jailbreaks require “adversarial prompting,” and noted the GAO used paid U.S. accounts. The companies also said the GAO tested models that have since been replaced. “To the companies’ credit, their safeguards have continued to evolve since we conducted these tests,” GAO director Kevin Walsh told reporters. “However, we conclude that, in its current state, this technology can be dangerous, and its misuse represents a growing risk to our nation.”
Only one company suspended the investigators’ accounts. Two others removed parts of the chats within days, and one of those accounts was still active less than two weeks later. Walsh declined to say whether any company alerted law enforcement, and called late suspensions “a trigger that may come too late.”
Why this matters
The report was requested by Senator Maggie Hassan, Democrat of New Hampshire, who is expected to chair the Senate Homeland Security and Governmental Affairs Committee next year, when Democrats take the majority. “If Congress doesn’t act, it is a question of when, not if, a bad actor will use AI to plan and carry out a major attack,” Hassan told reporters Tuesday. She warned safer new models don’t fix older ones still running the safeguards the GAO broke.
Hassan said AI safety legislation will be “one of our first areas of legislative focus” once she takes over the committee, starting with an early hearing on AI threats to homeland security and critical infrastructure. She also called for rules blocking the release of models “that don’t have adequate safety safeguards,” and a requirement that companies report illegal activity to law enforcement.
The report lands the same week OpenAI shipped GPT-6.2 Astra and Sol and Microsoft rolled out its own in-house models at Ignite. A tool that turns weeks of planning into hours changes who can plan an attack, and the four companies behind the most-used chatbots in America just failed the same test, on the same day, in front of the same investigators.
Congress has held plenty of AI hearings that ended in hand-wringing. This one starts with a government report, named companies, and a senator who will run the committee.
Frequently asked questions
What did the GAO report find about AI chatbots?
All four produced plans for mass-casualty bombings and grid attacks within 12 hours. The GAO called them largely viable.
Which AI chatbots were tested by the GAO?
ChatGPT, Claude, Gemini and Grok. The GAO did not say which chatbot produced which plan.
Did the chatbots refuse the requests at first?
Yes, at first. Investigators posing as regular users then got past the safeguards with common bypass techniques.
What happens next after the GAO report?
Senator Maggie Hassan plans an early hearing on AI threats to homeland security, and says AI safety legislation will be a first priority.
Sources: Associated Press, The Washington Sun, NBC News
