Nov 18, 2026 5:49 PM
Undercover investigators bypassed safety guardrails in less than 12 hours, manipulating popular systems to engineer credible blueprints for mass-casualty bombings and devastating strikes against the American electrical grid.
Four of the world's most widely used AI chatbots helped undercover U.S. government investigators develop potentially deadly attack plans in less than 12 hours, even after initially refusing to provide information that could be used to harm people or critical infrastructure.
In a newly released investigation, the Government Accountability Office (GAO) found that ChatGPT, Claude, Gemini, and Grok could be manipulated into providing information to help plan mass-casualty bombings and attacks on the U.S. electrical grid.
Investigators went beyond simply asking the chatbots dangerous questions. They checked the information against real-world conditions, visited potential locations, and consulted explosives and electrical infrastructure experts.
Ultimately, the GAO determined that all four systems produced information credible enough to help someone attempting an actual attack.
Related stories
“In each case, GAO’s prompts triggered the platforms’ safeguards, such as denial, removal of information, or suspending accounts,” investigators wrote. “GAO was able to circumvent those safeguards by using common safeguard-bypassing techniques.”
AI Safety Restrictions Failed in All Four Tests
In spring 2026, GAO investigators created undercover accounts and posed as ordinary users with basic technical knowledge of artificial intelligence.
The agency assigned two chatbots scenarios involving mass-casualty bombings and two others to attacks intended to disrupt electrical infrastructure.
Initially, all four systems refused requests for dangerous information. Some also removed conversations or suspended accounts after testing.
However, investigators eventually circumvented those restrictions using techniques designed to manipulate how AI systems interpret requests.
The resulting responses went far beyond general discussions of violence or infrastructure vulnerabilities.
In the bombing scenarios, the chatbots provided information to plan attacks on a hotel parking garage and a convention center, including specific dates, times, and methods. In the electrical-grid scenarios, the systems identified substations in major U.S. cities and generated plans involving drones or firearms.
The GAO withheld sensitive operational details and anonymized which chatbot produced each plan to prevent the report itself from becoming a resource for attackers.
Investigators Tested Whether the Plans Could Actually Work
One particularly revealing aspect of the investigation was the effort to determine whether the AI-generated information was realistic.
Rather than accepting chatbot responses at face value, investigators traveled to suggested locations, checked material availability, examined physical security conditions, and consulted outside specialists.
The results also exposed the technology's limitations.
One chatbot generated maps containing incorrectly positioned roads. In another case, investigators found that only two of four suggested electrical substations were physically vulnerable as the system described.
Nevertheless, the GAO concluded that the information from all four platforms was enough to help someone plan an attack.
A broad overview of the output provided by "Platform C" to GAO investigators during covert testing of AI systems ability to help plan an attack. (Image Source: GAO)The Bigger Threat May Be Speed, Not New Knowledge
Perhaps the most troubling finding was not that artificial intelligence could reveal previously secret information. Much of the material used to develop the plans was already available through public sources.
What changed was how quickly the information could be collected, organized, and turned into a potentially actionable plan.
The GAO found that tasks that might otherwise require weeks or months of research could be completed within hours.
For someone without specialized knowledge, a chatbot could reduce the expertise and effort needed to evaluate dangerous ideas, find relevant information, and develop a plan.
That raises a difficult question for AI developers: How effectively can a system distinguish between legitimate research and someone assembling publicly available information for malicious purposes?
An artistic rendering of an AI-generated map highlighting electrical substations, included in the GAO investigation into AI chatbots ability to assist in planning attacks. (Image Source: GAO)AI Companies Dispute Parts of the Findings
All four companies reviewed the GAO's findings and described efforts to strengthen their safety systems.
Anthropic questioned the study's small sample size and lack of a control group, arguing that the testing could not establish how much AI improved a user's ability to carry out dangerous tasks. The company also noted that it had expanded its safeguards since the tests.
OpenAI emphasized that ChatGPT refused numerous requests and argued that the report did not always clearly distinguish between information generated by the chatbot and material supplied by investigators or obtained from outside sources.
The GAO acknowledged that its testing was not statistically significant and represented a snapshot of the technology rather than an extensive comparison of the platforms.
Still, investigators maintained that the results presented a genuine security concern.
The report does not establish how frequently such safeguards fail or whether the same approaches would work against newer versions of the systems.
It does, however, provide a concrete example of how commercial AI tools can turn publicly accessible information into potentially dangerous plans.
As AI systems become more capable, preventing that kind of misuse may prove considerably harder than teaching a chatbot to refuse an obviously dangerous question.
“Generative AI is dramatically increasing productivity and decreasing the time it takes to perform many tasks,” the GAO report reads. “These advancements are enabling malicious actors to pursue nefarious actions faster and more efficiently and empowering less-skilled individuals by filling knowledge gaps through providing detailed expertise in sensitive areas.”
Tim McMillan is a retired law enforcement executive, investigative reporter and co-founder of The Debrief. His writing typically focuses on defense, national security, the Intelligence Community and topics related to psychology. You can follow Tim on Twitter: @LtTimMcMillan. Tim can be reached by email: [email protected] or through encrypted email: [email protected]
Comments
Powered by RoundtableAn AI/DeFi-powered Enterprise Media Operating System. Learn more at Roundtable.io.© Roundtable 2026. By using this site you agree to the Terms of Use and Privacy Policy


