From BigGo Finance · BigGo Finance · · 1m 10s
US GAO Finds AI Models Generate Terror Attack Plans
US GAO extracts actionable terror plans from ChatGPT, Gemini, and two other AI models within 12 hours.
In brief
US GAO extracts actionable terror plans from ChatGPT, Gemini, and two other AI models within 12 hours. US GAO tested four AI models and found investigators could extract plans to disable power grids or carry out explosive attacks within 12 hours. Originally reported by BigGo Finance.
US GAO Tests AI Models

US Government Accountability Office tested four generative AI services including ChatGPT, Gemini, Claude and Grok.
Investigators Extract Actionable Plans
Investigators succeeded within 12 hours in extracting plans detailed enough to disable power grids or carry out explosive attacks.
AI Safeguards Defeated
Each service's safeguards triggered but were defeated by routine circumvention techniques.
Actionable Terror Plans

Two services produced power grid disablement plans while the other two issued explosive attack instructions.
Expert Assessment
GAO assessed most plans as actionable based on expert review and on-the-ground conditions.
Jailbreaking Technique
Known as 'jailbreaking', the method evades AI restrictions to elicit prohibited answers.
The short version
US GAO tested four AI models and found investigators could extract plans to disable power grids or carry out explosive attacks within 12 hours.
Read the TL;DR — this story on one card.





