Government Test Finds AI Chatbots Produced Viable Attack Plans
Government Accountability Office investigators posing as ordinary users got four leading AI chatbots to produce terrorist attack plans. The report was released Wednesday.
- 12 hours or less: time each platform took to generate a plan.
- Two chatbots planned vehicle bombings at a hotel garage and convention center.
- Two others targeted electrical substations, one by drone, one by gunfire.
- Each chatbot refused at first before investigators bypassed its safeguards.
Why it matters: The GAO concluded the plans were largely viable. A bomb squad commander said both bomb recipes would explode.
- The GAO said the details were not new but AI compressed weeks of planning into hours.
- Investigators also found errors, including wrong road placements and substations behind high walls.
How 29 sources split on this story
Where they split: The dispute is whether the tests showed chatbots added real help beyond what a web search provides.
Center coverage, 11 sources: The center presents the GAO findings alongside detailed company pushback on the testing methods.
Left coverage, 9 sources: The left frames the findings as proof that AI safeguards fail and that Congress must regulate the industry.
What’s next: Hassan wants rules blocking release of models without adequate safeguards. She also wants mandatory reporting of illegal activity to law enforcement.
- Companies say newer models are safer; Hassan notes some tested models remain available.
- How much did the chatbots speed up planning compared with ordinary internet searches?
- Did any test trigger a notification to law enforcement?


















