Congress's auditors asked four chatbots to plan an attack. All four did, in under 12 hours
GAO investigators got Claude, Grok, ChatGPT and Gemini past their guardrails and walked away with targets, dates and shopping lists for a bombing and a grid attack.
Last spring, investigators from the Government Accountability Office opened accounts on the four most popular chatbots in America and asked them for help killing people. They wanted two things. One was a plan to set off an improvised explosive device built for mass casualties. The other was a plan to hit the electrical grid hard enough to cause economic damage.
They got both. The report, GAO-27-108814, went public on Wednesday, and the auditors sum up its central finding in one line: "Each platform took less than 12 hours to generate a plan for the attacks."
The four platforms were Claude, Grok, ChatGPT and Gemini, picked because usage data from April 2026 showed they were the most used generative AI tools in the country. GAO hid which was which. In the report they appear as Platforms A through D, assigned at random, two to each scenario. You can read the whole report and still not know whose model suggested the parking garage.
What the plans contained
Platform A produced the most chilling output. According to the report, it picked an exact date, an exact time and a specific hotel parking garage in a major American city, then recommended a bomb hidden in a vehicle. It explained how to keep the vehicle from being tracked and named stores along the route where the supplies could be bought. Platform B also gave an exact date and time, with a specific convention center as the target. ABC News reported that one of the timelines recommended setting up a shell company and spreading purchases across several states.
The grid plans were blunter. Platform C proposed a drone attack on named substations in a major city. Platform D proposed a firearms attack on substations and suggested a season of the year to carry it out. ABC said both explained how to pick out and acquire the drones and guns involved.
GAO did not take the chatbots' word for any of this. Investigators visited stores to see whether the materials were on sale, checked sites, ran surveillance and brought in outside experts, including what ABC described as "a bomb squad commander at a 'major metropolitan area' fire department." They built nothing. Their verdict on all four plans was that the information was sufficient for a user to plan and carry out an attack.
The guardrails held for a while
Every platform pushed back at first. The report says all four activated their safety guardrails, and in every case the tester got past them. GAO's summary says it "was able to circumvent those safeguards by using common safeguard-bypassing techniques." The testers worked under undercover personas, and what they needed most was patience and a believable story.
GAO doesn't publish the prompts, which is sensible. The auditors make the broader point themselves: "Every safeguard used in the industry is vulnerable to jailbreaks," and "there are no current generative AI systems that are completely immune to such misuse."
The companies noticed some of it. A footnote records that "The company associated with Platform B suspended the user account after the testing period concluded." By then the plan was already in the investigators' hands. Two other platforms deleted parts of the chat history the day after testing.
Kevin Walsh, a GAO director, told ABC News that "in its current state, this technology can be dangerous, and its misuse represents a growing risk to our nation."
The companies' replies follow a familiar script. OpenAI told ABC it takes the findings seriously and that the testing happened "months ago on an older version of ChatGPT." Google, in a statement quoted by Newsmax and carried on Yahoo News, said "We recognize our work on the safety and security of our AI products is never complete," and also pointed out that the investigators used older models. SpaceXAI, which runs Grok, said: "We appreciate that the report acknowledges that pre-written prompts failed on all four platforms." Anthropic pushed back on the method. Per the same Newsmax report, it said "We do not dispute that a determined tester can sometimes get harmful content out of an AI system," then argued the test was too small to be statistically meaningful and that it couldn't reproduce the results without the investigators' conversations.
Some of that is fair. GAO itself says its results are not a statistically valid sample, and there was no control group. Four platforms, two scenarios, one round of testing. Every model tested has since been replaced by a newer one. Still, "that was the old version" is an odd defense from companies that ship a new model every few months, each more capable than the last. Whatever you are using today will be the old version by the time an auditor gets to it.
Sen. Maggie Hassan of New Hampshire asked for the report. In a statement on Wednesday she said "The AI public safety threat is no longer theoretical," and she told ABC that "These companies are facing extraordinary competitive pressures to get their models out the door." She plans hearings on AI safety next year and wants legislation that would force stronger safeguards before models are released.
GAO made no formal policy recommendations. It did not say whether any of its tests set off an alert at a law enforcement agency, and nothing in the report suggests one did. Each platform needed less than half a working day to hand over a target, a date and a shopping list, and the strongest response the report records from any company is one account suspended after the job was done.
Sources: GAO: Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (GAO-27-108814), ABC News, Fox News, Newsmax via Yahoo News.
More from the feed
OpenAI rates its new model a self-improvement risk. UK testers jailbroke it faster
GPT-6.2 Astra is the first model OpenAI has shipped at its High self-improvement level. The same system card says Britain's AI Security Institute broke its safeguards in half the queries it needed last time.
An AI hacking tool went closed-source. The FBI says the old version is everywhere
Artex has been tied to break-ins at ConnectWise, Barracuda and South Korean banks. Its developer locked it down, but the FBI says the open-source copy remains 'widely accessible'.
Someone built a bot that turns security patches into attacks. Google found its server
Google's threat researchers stumbled on an exposed machine running exploit_agent, a framework that reads fresh patches, writes the exploit and fires it with nobody at the keyboard.