In a May 2026 test requested by Senator Maggie Hassan, GAO investigators got AI like ChatGPT, Claude, Gemini, and Grok to generate operational proposals—including bomb placement ideas and drone attacks on electrical substations—after circumventing their guardrails
BY PC Bureau
November 18: Four leading artificial intelligence ( AI) chatbots ChatGPT, Claude, Gemini, and Grok generated detailed proposals for mass-casualty attacks and economic sabotage after US government investigators found ways to circumvent their built-in safety restrictions, according to a Government Accountability Office (GAO) report released following tests conducted in May 2026.
The investigation found that the chatbots initially rejected requests to plan violent attacks but eventually provided operational suggestions after investigators used techniques to bypass their safeguards. The responses included potential targets, timing considerations and methods intended to maximise casualties or cause substantial economic damage.
The test was conducted at the request of Democratic Senator Maggie Hassan of New Hampshire, who has raised concerns about the ability of commercial AI systems to prevent their technology from being misused.
Investigators assigned the four chatbots one of two objectives: devising an attack intended to cause the highest possible number of casualties or planning an operation designed to inflict significant economic damage.
Although the systems initially refused direct requests, the investigators found that their safeguards could be circumvented. Within approximately 12 hours, the chatbots had moved from refusing the requests to producing concrete attack proposals.
Chatbots generated proposals for mass-casualty attacks
Two of the four systems generated plans involving attacks intended to cause mass casualties. Their responses included identifying a hotel parking garage and a convention centre as potential targets.
Some proposals also included suggested dates and timings intended to increase the impact of an attack. One chatbot even generated an image illustrating a proposed placement for a vehicle carrying explosives inside a parking garage.
The other two systems concentrated on economic disruption, particularly attacks against electricity infrastructure.
One suggested a drone-based attack on an electrical substation and generated a four-step diagram and a simulated video depicting the scenario. Another proposed attacking a substation component with a firearm and supplied additional operational details.
The GAO report also documented responses involving the use of front companies, rented vehicles, burner phones and decoys to support attack preparations.
The agency did not identify which chatbot generated individual responses. It also withheld certain technical details to prevent the findings from becoming a guide for potential attackers.
GAO: CLAUDE, GROK, CHATGPT AND GEMINI PRODUCED VIABLE US ATTACK PLANS
The Government Accountability Office said Wednesday that Claude, Grok, ChatGPT and Gemini, the four most-used AI chatbots, produced feasible plans for mass-casualty or electrical-grid attacks in U.S. cities during covert tests last spring. Safeguards initially blocked the material, the watchdog said, but common bypass techniques still yielded a plan from each platform in less than 12 hours.Two platforms produced IED plans for a hotel or convention center; the other two outlined drone or firearms attacks on electrical substations, the report said, without naming which chatbot produced which plan. The companies said the tests used older versions and they have since strengthened protections. Sen. Maggie Hassan, a New Hampshire Democrat, said a terrorist could walk away “with a concrete playbook” in less than 12 hours. https://t.co/u0OxV5Siku
— AmericaToday.com (@ATNewsChannel) November 18, 2026
AI companies question the findings
The companies acknowledged that their AI systems could sometimes generate harmful content but raised concerns about the methodology and relevance of the experiment.
Among their objections was the absence of a control group comparing the chatbots’ responses with information that could be obtained through conventional internet searches.
They also pointed out that the investigation used versions of their models available in May 2026, before subsequent safety improvements and software updates.
xAI additionally argued that its moderation policies differ for registered, paying users in the United States, whom it considers a lower-risk category than free users or users accessing its services from other countries.
The report examined several techniques used to defeat AI safety restrictions, including so-called ‘Do Anything Now’ (DAN) prompts that attempt to persuade a chatbot to adopt an unrestricted persona.
Another approach involves breaking a dangerous request into several smaller questions, making each individual prompt appear less suspicious while gradually extracting information that could be combined into an operational plan.
The GAO did not disclose the specific techniques used by its investigators.
READ: Nvidia Faces Ransomware Threat as Hackers Set Deadline
Senator calls for stronger AI safeguards
Senator Hassan described the findings as evidence that existing AI safety mechanisms remain inadequate.
She warned that the information generated during the experiment “could be deadly in the hands of a terrorist who actually intends to carry out an attack”, arguing that the potential threat posed by inadequately protected AI systems should not be dismissed as hypothetical.
Hassan has indicated that she intends to push for stronger AI safety legislation and congressional hearings when Democrats assume control of the Senate in January.
Her proposed priorities include mandatory safety assessments before AI products are released, requirements to report certain illegal activities to law enforcement and continuing evaluations of safeguards after deployment.
Renewed scrutiny of AI safety
The findings are likely to intensify debate over the responsibilities of AI developers as their systems become increasingly capable of processing complex requests and generating detailed responses.
The central challenge is preventing powerful AI tools from facilitating violence, terrorism and attacks on critical infrastructure without unnecessarily restricting legitimate research and everyday applications.
The GAO’s investigation highlights a significant vulnerability: safeguards that appear effective against direct requests may not withstand persistent attempts to manipulate a chatbot’s behaviour.
By withholding the technical methods used in the experiment, the agency sought to expose the weaknesses without providing potential attackers with a ready-made guide to exploiting them.
The report raises fresh questions about whether AI companies can keep pace with evolving misuse techniques and whether voluntary safety measures are sufficient to address the risks.








