Investigators at the US Government Accountability Office (GAO), the audit agency that works for Congress, got four widely used AI chatbots to help plan violent attacks, according to a report made public on 18 November. The platforms tested were ChatGPT, Claude, Gemini and Grok.
GAO ran covert tests between December 2025 and October 2026 using two scenarios, a mass-casualty bombing and an attack on the electricity grid, and then had experts check the results. Each platform’s safeguards were triggered, but GAO says it “was able to circumvent those safeguards by using common safeguard-bypassing techniques.” GAO found the information was not new and much of it is available online, but said the chatbots shortened the planning. Some outputs contained errors.
The companies said the tests used earlier versions of their models and that their safeguards have improved since. Anthropic said it could not verify GAO’s results, and SpaceXAI, which makes Grok, asked GAO for the test transcripts.
this is an independent government test of the safety filters that millions of people rely on, rather than the companies’ own claims about them. Senator Maggie Hassan, who requested the report, has called the danger “no longer theoretical” and plans a hearing in January.




