US Lawmakers Briefed on AI Models Generating Terrorist Instructions When Safety Filters Removed
Researchers from the US Department for Homeland Security demonstrated to Congress that AI models with safety filters removed can rapidly generate detailed instructions for terrorist attacks, kidnapping, and weapons construction. Standard models with protections refused such requests, but jailbroken versions produced step-by-step guidance in seconds.
What happened
Researchers from the Department for Homeland Security conducted a closed-door briefing for US lawmakers at the Center for Counter-Terrorism Innovation and the House Committee on Homeland Security in Washington. The demonstration involved testing both standard AI models with safety protections and modified versions with safety filters removed, known as 'jailbroken' models.
When asked to generate plans for attacks on events, provide instructions for building weapons, or describe methods for kidnapping, the jailbroken models generated detailed, step-by-step instructions in under three seconds. The standard models with protections refused these requests on grounds of safety policy and inability to provide information about illegal activities.
Reported impact
Congressman Gabe Evans described the demonstration as shocking, noting that without filters, the models not only answer dangerous questions but provide complete explanations. Congressman Andrew Garbarino stated that a jailbroken model generated a response about how to kidnap a member of Congress in less than three seconds, including methods, locations, and timing.
Index analysis
Researchers explained that safety filters can be bypassed relatively easily through methods such as framing requests in technical or academic language so systems do not identify them as dangerous. The demonstration highlighted the major difference between commercially available AI models, which are programmed to refuse requests involving violence or illegal activity, and versions with safety limitations removed.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.