A government investigation has found that American artificial intelligence models can be manipulated to assist in planning mass-casualty attacks. The Government Accountability Office conducted the audit and released its findings this week. Researchers involved in the probe successfully prompted chatbots to produce dangerous content.
The study focused on widely used commercial AI systems developed by U.S. companies. Investigators tested whether these tools could be steered toward harmful outputs through targeted prompts. The results showed that several models failed to block requests for attack-related information.
Researchers used carefully crafted questions to bypass safety filters built into the chatbots. Once past those barriers, the systems generated detailed data applicable to planning violent acts. The GAO described the outputs as material that could support mass-casualty attack planning.
The findings raise concerns about the safeguards companies place on publicly available AI tools. Current safety measures rely heavily on content filters and refusal training. The audit suggests these defenses can be circumvented with relative ease.
Lawmakers have begun reviewing the report and its implications for national security. The GAO recommended that federal agencies strengthen oversight of AI safety testing. It also urged closer coordination between developers and government security experts.
AI companies have defended their safety efforts while acknowledging the challenges involved. Several firms said they continuously update their models to address emerging risks. Independent testing remains a key part of their development process.
The report adds to a growing debate over how to regulate powerful AI systems. Officials must balance innovation with the need to prevent misuse. The GAO plans to monitor how agencies respond to its recommendations.




