WHAT YOU NEED TO KNOW
  • GAO investigators bypassed safeguards in ChatGPT, Claude, Gemini and Grok during testing conducted in May 2026.
  • Within 12 hours, chatbots produced two mass casualty plans and two proposals targeting the electrical grid.
  • The plans included explosives guidance, attack timing, diagrams, ammunition recommendations, decoy vehicles and a simulated drone attack video.
  • Sen. Maggie Hassan wants stronger safety requirements, regular safeguard updates, law enforcement reporting and a Senate hearing on AI safety.

Several popular artificial intelligence chatbots provided detailed plans for mass casualty attacks and strikes against the electrical grid after Government Accountability Office investigators bypassed their safety systems, according to a watchdog report released Wednesday morning.

Within 12 hours, the congressional watchdog obtained two plans intended to cause mass casualties and two additional plans designed to inflict enormous economic damage. The chilling results emerged from testing conducted in May 2026.

Investigators tested ChatGPT, Claude, Gemini and Grok, randomly assigning each chatbot to develop an attack that would either maximize casualties or economic damage. One proposed bombing a hotel parking garage, while another selected a convention center.

The other chatbots produced plans targeting electrical substations. One recommended using a drone, while another instructed investigators to shoot a substation component with a rifle.

Every chatbot initially refused to help plan the attacks. GAO investigators nevertheless found ways around those safeguards, though the report withheld the methods to avoid handing useful instructions to potential copycats.

The report noted several established techniques for circumventing chatbot restrictions. The “Do Anything Now” prompt directs a chatbot to roleplay as an alter ego unconstrained by safety rules, while other approaches disguise malicious requests or gather information incrementally.

GAO also declined to identify which chatbot produced each answer. What the investigators received, however, went far beyond vague or generic information and included specific operational details.

The two proposals involving improvised explosive devices suggested particular times and dates intended to maximize damage. One chatbot generated an image showing where an explosive should be placed inside a parking garage.

Should Election Day be a federal holiday?

By completing the poll, you agree to receive emails from Left Scoop News, occasional offers from our partners and that you've read and agree to our privacy policy and legal statement.

Another chatbot recommended starting a shell company, renting a catering vehicle, leasing farmland, purchasing a burner phone and sending decoy trucks toward the target. It also supplied advice about creating the explosive.

The proposed drone attack against an electrical substation featured a four step diagram and what the report described as “a realistic, simulated video of a drone attack.” The rifle proposal identified specific ammunition and selected a location near a large international airport and multiple data centers to increase economic damage.

Sen. Maggie Hassan of New Hampshire, a Democrat who is in line to become chair of the Senate Homeland Security and Governmental Affairs Committee, requested the test. She said she plans to pursue stricter safeguards and hold at least one hearing on AI safety early next year.

“The same information that testers were able to get a hold of could be deadly in the hands of a terrorist who actually intends to carry out an attack,” Hassan said on a conference call Tuesday. “The current safeguards that AI platforms have in place really don’t go far enough, and the threat that AI poses to public safety is not theoretical. 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.”

The companies behind the chatbots acknowledged that users could obtain dangerous information, but they challenged parts of the report. They argued that the experiment lacked a control group showing whether the same material was available through search engines or whether chatbots saved investigators time.

The companies also emphasized that testing occurred in May 2026 and said safety improvements had been made since then. SpaceXAI, which makes Grok, noted that GAO conducted its test through a paid account.

“SpaceXAI provides unique moderation policies for registered, paid US account holders that reflect the lower risk profile of such accounts compared to free or foreign accounts,” the company said, according to the report. GAO would not disclose whether its activity triggered any notification to law enforcement.

In one instance, a chatbot blocked access to a GAO account only after providing the information. Another retroactively removed parts of a conversation about electrical grid vulnerabilities, although the report said it was “unclear whether this action was initiated as part of a guardrail or if by error.”

Hassan wants rules preventing companies from releasing chatbot models without adequate safety protocols. She also called for companies to report illegal activity to law enforcement and establish “a constant mechanism” for regularly updating safeguards.

“These companies are facing extraordinary competitive pressures to get their models out the door and to the public as quickly as possible,” Hassan told reporters. “And one of the things that will be necessary going forward to make sure that they’re all dealing with a level playing field when it comes to safety is making sure that before models are released, they have safeguards that can’t be breached in this way.”