AI Chatbots Suggested How To Bomb Public Places and Damage Power Stations in Government Test

Investigators obtained bomb and power-grid attack information from ChatGPT, Claude, Gemini and Grok by using common jailbreak techniques, according to a govreport

ai platform
AI chatbots provided detailed plans for bomb attacks and power station sabotage based on a US government test Salvador Rios/Unsplash

Believe it or not, artificial intelligence (AI) chatbots can now provide instructions on how to carry out bomb attacks in public places during a government test in the United States (US), after investigators found ways to bypass safety rules.

Investigators from the Government Accountability Office (GAO), a US congressional watchdog agency, obtained detailed attack plans from four popular AI platforms, particularly ChatGPT, Claude, Gemini, and Grok, within 12 hours, according to a report released on November 18, Wednesday.

The GAO conducted the test in spring 2026 at the request of Senator Maggie Hassan, a Democrat from New Hampshire, who is set to lead the Senate Homeland Security and Governmental Affairs Committee.

The four AI chatbots were each given scenarios involving either a bomb attack or an attempt to damage the country's electrical grid, and while all platforms initially refused to assist with the requests, investigators eventually found loopholes to go around their restrictions.

The investigators initially received refusals from all four platforms but eventually bypassed their safeguards using common techniques known as jailbreaks. These include roleplaying, disguising harmful requests through language, and gradually obtaining restricted information through seemingly harmless prompts.

However, the report did not identify which chatbot provided each response to prevent others from copying the methods used in the experiment.

Chatbots Generate Attack Plans With Dates, Locations, and Ways

Two of the AI platforms generated plans for improvised explosive device attacks targeting a hotel parking garage and a convention centre in major US cities.

Their responses were down to the tiniest details; they planned out the possible dates and times for the attacks, as well as information on how to prepare them, which was alarming.

One chatbot also provided an image showing where the bomb could be planted in the parking garage, while another laid out a timeline that found ways on how to not get caught by evading law enforcement.

On the other hand, the other two platforms provided plans for targeting electrical substations. After the prompt, there were also other parts of their responses, where one chatbot generated a diagram and a simulated video that shows a drone attack on a power station.

Another suggested using a firearm against a substation component and identified a location near an international airport and several data centres.

The GAO verified the feasibility of the information through online research, surveillance, site visits, and consultations with a bomb squad commander and an electric industry expert.

GAO Findings Renew Calls for Stricter AI Regulations

The findings have renewed calls for stronger regulations on AI platforms, with Senator Hassan pushing for congressional hearings and legislation to address the risks identified in the tests.

Hassan warned that information generated during the test could become dangerous if accessed by citizens with the ill-intent or was hired to carry out actual, real-life attacks.

'The AI public safety threat is no longer theoretical. If Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack,' she said in a US conference.

As an accountability check, the companies behind the four platforms acknowledged that their systems could produce harmful information.

In a statement to ABC News, OpenAI said it took the findings seriously, but then emphasised that the government test used an older version of ChatGPT rather than the most recent model.

The company added that it had since introduced safety improvements across its GPT-6 family of models to strengthen safeguards against harmful uses.

According to the Associated Press, Google also recognised that the GAO tested earlier versions of its models, while SpaceXAI said it 'continues to work diligently' to improve the safety of its models with the upcoming new releases.